Initial implementation of compaction MCP server

Ephemeral, per-session context management for LLMs with 8 MCP tools: store, query, status, compact, pin, forget, configure, update. Features: scoring with decay/importance/access, Jaccard dedup, summary promotion, budget-based eviction, auto-detection of client context window via MCP hooks.
2026-06-05 23:14:02 +00:00 · 2026-03-07 16:57:12 +00:00
commit 0ddd0e4598
8 changed files with 1050 additions and 0 deletions
@@ -0,0 +1 @@
+compaction-mcp
@@ -0,0 +1,17 @@
+module compaction-mcp
+
+go 1.26.1
+
+require github.com/mark3labs/mcp-go v0.45.0
+
+require (
+	github.com/bahlo/generic-list-go v0.2.0 // indirect
+	github.com/buger/jsonparser v1.1.1 // indirect
+	github.com/google/uuid v1.6.0 // indirect
+	github.com/invopop/jsonschema v0.13.0 // indirect
+	github.com/mailru/easyjson v0.7.7 // indirect
+	github.com/spf13/cast v1.7.1 // indirect
+	github.com/wk8/go-ordered-map/v2 v2.1.8 // indirect
+	github.com/yosida95/uritemplate/v3 v3.0.2 // indirect
+	gopkg.in/yaml.v3 v3.0.1 // indirect
+)
@@ -0,0 +1,39 @@
+github.com/bahlo/generic-list-go v0.2.0 h1:5sz/EEAK+ls5wF+NeqDpk5+iNdMDXrh3z3nPnH1Wvgk=
+github.com/bahlo/generic-list-go v0.2.0/go.mod h1:2KvAjgMlE5NNynlg/5iLrrCCZ2+5xWbdbCW3pNTGyYg=
+github.com/buger/jsonparser v1.1.1 h1:2PnMjfWD7wBILjqQbt530v576A/cAbQvEW9gGIpYMUs=
+github.com/buger/jsonparser v1.1.1/go.mod h1:6RYKKt7H4d4+iWqouImQ9R2FZql3VbhNgx27UK13J/0=
+github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
+github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/frankban/quicktest v1.14.6 h1:7Xjx+VpznH+oBnejlPUj8oUpdxnVs4f8XU8WnHkI4W8=
+github.com/frankban/quicktest v1.14.6/go.mod h1:4ptaffx2x8+WTWXmUCuVU6aPUX1/Mz7zb5vbUoiM6w0=
+github.com/google/go-cmp v0.5.9 h1:O2Tfq5qg4qc4AmwVlvv0oLiVAGB7enBSJ2x2DqQFi38=
+github.com/google/go-cmp v0.5.9/go.mod h1:17dUlkBOakJ0+DkrSSNjCkIjxS6bF9zb3elmeNGIjoY=
+github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
+github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
+github.com/invopop/jsonschema v0.13.0 h1:KvpoAJWEjR3uD9Kbm2HWJmqsEaHt8lBUpd0qHcIi21E=
+github.com/invopop/jsonschema v0.13.0/go.mod h1:ffZ5Km5SWWRAIN6wbDXItl95euhFz2uON45H2qjYt+0=
+github.com/josharian/intern v1.0.0/go.mod h1:5DoeVV0s6jJacbCEi61lwdGj/aVlrQvzHFFd8Hwg//Y=
+github.com/kr/pretty v0.3.1 h1:flRD4NNwYAUpkphVc1HcthR4KEIFJ65n8Mw5qdRn3LE=
+github.com/kr/pretty v0.3.1/go.mod h1:hoEshYVHaxMs3cyo3Yncou5ZscifuDolrwPKZanG3xk=
+github.com/kr/text v0.2.0 h1:5Nx0Ya0ZqY2ygV366QzturHI13Jq95ApcVaJBhpS+AY=
+github.com/kr/text v0.2.0/go.mod h1:eLer722TekiGuMkidMxC/pM04lWEeraHUUmBw8l2grE=
+github.com/mailru/easyjson v0.7.7 h1:UGYAvKxe3sBsEDzO8ZeWOSlIQfWFlxbzLZe7hwFURr0=
+github.com/mailru/easyjson v0.7.7/go.mod h1:xzfreul335JAWq5oZzymOObrkdz5UnU4kGfJJLY9Nlc=
+github.com/mark3labs/mcp-go v0.45.0 h1:s0S8qR/9fWaQ3pHxz7pm1uQ0DrswoSnRIxKIjbiQtkc=
+github.com/mark3labs/mcp-go v0.45.0/go.mod h1:YnJfOL382MIWDx1kMY+2zsRHU/q78dBg9aFb8W6Thdw=
+github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
+github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
+github.com/rogpeppe/go-internal v1.9.0 h1:73kH8U+JUqXU8lRuOHeVHaa/SZPifC7BkcraZVejAe8=
+github.com/rogpeppe/go-internal v1.9.0/go.mod h1:WtVeX8xhTBvf0smdhujwtBcq4Qrzq/fJaraNFVN+nFs=
+github.com/spf13/cast v1.7.1 h1:cuNEagBQEHWN1FnbGEjCXL2szYEXqfJPbP2HNUaca9Y=
+github.com/spf13/cast v1.7.1/go.mod h1:ancEpBxwJDODSW/UG4rDrAqiKolqNNh2DX3mk86cAdo=
+github.com/stretchr/testify v1.9.0 h1:HtqpIVDClZ4nwg75+f6Lvsy/wHu+3BoSGCbBAcpTsTg=
+github.com/stretchr/testify v1.9.0/go.mod h1:r2ic/lqez/lEtzL7wO/rwa5dbSLXVDPFyf8C91i36aY=
+github.com/wk8/go-ordered-map/v2 v2.1.8 h1:5h/BUHu93oj4gIdvHHHGsScSTMijfx5PeYkE/fJgbpc=
+github.com/wk8/go-ordered-map/v2 v2.1.8/go.mod h1:5nJHM5DyteebpVlHnWMV0rPz6Zp7+xBAnxjb1X5vnTw=
+github.com/yosida95/uritemplate/v3 v3.0.2 h1:Ed3Oyj9yrmi9087+NczuL5BwkIc4wvTb5zIM+UJPGz4=
+github.com/yosida95/uritemplate/v3 v3.0.2/go.mod h1:ILOh0sOhIJR3+L/8afwt/kE++YT040gmv5BQTMR2HP4=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
+gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
@@ -0,0 +1,70 @@
+package main
+
+import (
+	"context"
+	"flag"
+	"fmt"
+	"os"
+	"strings"
+
+	"github.com/mark3labs/mcp-go/mcp"
+	"github.com/mark3labs/mcp-go/server"
+)
+
+const serverInstructions = `Context compactor — manages working memory within a token budget.
+
+At session start, call 'configure' with token_budget set to ~40% of your context window.
+Example: 200K context window → token_budget = 80000.
+
+Workflow:
+- 'store' important context (always include a summary for efficient compaction)
+- 'query' to retrieve stored information instead of re-reading sources
+- 'status' to check budget usage
+- 'compact' when usage is high — it frees space and identifies items needing summarization
+- 'update' to add summaries to items flagged by compaction`
+
+func main() {
+	budget := flag.Int("budget", 100000, "Token budget for context storage")
+	flag.Parse()
+
+	budgetExplicit := false
+	flag.Visit(func(f *flag.Flag) {
+		if f.Name == "budget" {
+			budgetExplicit = true
+		}
+	})
+
+	store := NewStore(*budget)
+
+	hooks := &server.Hooks{}
+	if !budgetExplicit {
+		hooks.OnAfterInitialize = append(hooks.OnAfterInitialize,
+			func(ctx context.Context, id any, req *mcp.InitializeRequest, res *mcp.InitializeResult) {
+				name := strings.ToLower(req.Params.ClientInfo.Name)
+				switch {
+				case strings.Contains(name, "claude"):
+					// Claude models: 200K context → 40% = 80K budget
+					store.Configure(80000, nil, 0)
+				case strings.Contains(name, "cursor"):
+					store.Configure(60000, nil, 0)
+				}
+			},
+		)
+	}
+
+	s := server.NewMCPServer(
+		"compactor",
+		"0.1.0",
+		server.WithToolCapabilities(false),
+		server.WithRecovery(),
+		server.WithInstructions(serverInstructions),
+		server.WithHooks(hooks),
+	)
+
+	registerTools(s, store)
+
+	if err := server.ServeStdio(s); err != nil {
+		fmt.Fprintf(os.Stderr, "server error: %v\n", err)
+		os.Exit(1)
+	}
+}
@@ -0,0 +1,446 @@
+package main
+
+import (
+	"crypto/rand"
+	"fmt"
+	"math"
+	"sort"
+	"strings"
+	"sync"
+	"time"
+)
+
+type Item struct {
+	CreatedAt   time.Time
+	AccessedAt  time.Time
+	Content     string
+	Summary     string
+	ID          string
+	Tags        []string
+	Tokens      int
+	AccessCount int
+	Importance  int
+	Pinned      bool
+}
+
+type CompactResult struct {
+	NeedsSummary []*Item
+	TokensFreed  int
+	TokensBefore int
+	TokensAfter  int
+	Evicted      int
+	Summarized   int
+	Deduplicated int
+}
+
+type Store struct {
+	items                map[string]*Item
+	tokenBudget          int
+	usedTokens           int
+	autoCompactThreshold float64
+	mu                   sync.Mutex
+	autoCompact          bool
+}
+
+func NewStore(tokenBudget int) *Store {
+	return &Store{
+		items:                make(map[string]*Item),
+		tokenBudget:          tokenBudget,
+		autoCompact:          true,
+		autoCompactThreshold: 0.9,
+	}
+}
+
+func (s *Store) Add(content, summary string, tags []string, importance int) *Item {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	if importance < 1 {
+		importance = 5
+	} else if importance > 10 {
+		importance = 10
+	}
+
+	tokens := EstimateTokens(content)
+	item := &Item{
+		ID:         newID(),
+		Content:    content,
+		Summary:    summary,
+		Tags:       tags,
+		Importance: importance,
+		CreatedAt:  time.Now(),
+		AccessedAt: time.Now(),
+		Tokens:     tokens,
+	}
+
+	s.items[item.ID] = item
+	s.usedTokens += tokens
+
+	if s.autoCompact && s.tokenBudget > 0 {
+		if float64(s.usedTokens)/float64(s.tokenBudget) > s.autoCompactThreshold {
+			s.compactLocked(0.8)
+		}
+	}
+
+	return item
+}
+
+func (s *Store) Get(id string) (*Item, bool) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	item, ok := s.items[id]
+	if ok {
+		item.AccessedAt = time.Now()
+		item.AccessCount++
+	}
+	return item, ok
+}
+
+func (s *Store) Remove(id string) bool {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	item, ok := s.items[id]
+	if !ok {
+		return false
+	}
+	s.usedTokens -= item.Tokens
+	delete(s.items, id)
+	return true
+}
+
+func (s *Store) Pin(id string) bool {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	item, ok := s.items[id]
+	if ok {
+		item.Pinned = true
+	}
+	return ok
+}
+
+func (s *Store) UpdateSummary(id, summary string) bool {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	item, ok := s.items[id]
+	if ok {
+		item.Summary = summary
+	}
+	return ok
+}
+
+func (s *Store) Query(query string, tags []string, limit int) []*Item {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	if limit <= 0 {
+		limit = 10
+	}
+
+	queryWords := wordSet(query)
+	var results []*Item
+
+	for _, item := range s.items {
+		if len(tags) > 0 && !hasAnyTag(item, tags) {
+			continue
+		}
+		item.AccessedAt = time.Now()
+		item.AccessCount++
+		results = append(results, item)
+	}
+
+	sort.Slice(results, func(i, j int) bool {
+		return s.queryScore(results[i], queryWords) > s.queryScore(results[j], queryWords)
+	})
+
+	if len(results) > limit {
+		results = results[:limit]
+	}
+	return results
+}
+
+func (s *Store) scoreLocked(item *Item) float64 {
+	if item.Pinned {
+		return math.MaxFloat64
+	}
+
+	age := time.Since(item.AccessedAt).Minutes()
+	recency := math.Exp(-age / 120.0) // half-life ~2 hours
+
+	importance := float64(item.Importance) / 10.0
+	access := math.Log1p(float64(item.AccessCount)) / 5.0
+
+	var sizePenalty float64
+	if s.tokenBudget > 0 {
+		sizePenalty = float64(item.Tokens) / float64(s.tokenBudget)
+	}
+
+	return (0.4 * importance) + (0.3 * recency) + (0.2 * access) - (0.1 * sizePenalty)
+}
+
+func (s *Store) queryScore(item *Item, queryWords map[string]struct{}) float64 {
+	base := s.scoreLocked(item)
+	if len(queryWords) == 0 {
+		return base
+	}
+
+	contentWords := wordSet(item.Content)
+	if item.Summary != "" {
+		for w := range wordSet(item.Summary) {
+			contentWords[w] = struct{}{}
+		}
+	}
+	for _, tag := range item.Tags {
+		contentWords[strings.ToLower(tag)] = struct{}{}
+	}
+
+	return base + (0.5 * jaccardSimilarity(queryWords, contentWords))
+}
+
+func (s *Store) Status() (budget, used, count int, usage float64) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	budget = s.tokenBudget
+	used = s.usedTokens
+	count = len(s.items)
+	if budget > 0 {
+		usage = float64(used) / float64(budget)
+	}
+	return
+}
+
+func (s *Store) BudgetTight() bool {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	if s.tokenBudget <= 0 {
+		return false
+	}
+	return float64(s.usedTokens)/float64(s.tokenBudget) > 0.8
+}
+
+func (s *Store) Configure(budget int, autoCompact *bool, threshold float64) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	if budget > 0 {
+		s.tokenBudget = budget
+	}
+	if autoCompact != nil {
+		s.autoCompact = *autoCompact
+	}
+	if threshold > 0 && threshold <= 1.0 {
+		s.autoCompactThreshold = threshold
+	}
+}
+
+func (s *Store) AutoCompact() bool {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.autoCompact
+}
+
+func (s *Store) AutoCompactThreshold() float64 {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.autoCompactThreshold
+}
+
+func (s *Store) Compact(targetUsage float64) CompactResult {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.compactLocked(targetUsage)
+}
+
+func (s *Store) compactLocked(targetUsage float64) CompactResult {
+	result := CompactResult{TokensBefore: s.usedTokens}
+
+	if s.tokenBudget <= 0 {
+		result.TokensAfter = s.usedTokens
+		return result
+	}
+
+	targetTokens := int(float64(s.tokenBudget) * targetUsage)
+	if s.usedTokens <= targetTokens {
+		result.TokensAfter = s.usedTokens
+		return result
+	}
+
+	// Phase 1: Summary promotion — replace content with summary to save tokens
+	for _, item := range s.items {
+		if s.usedTokens <= targetTokens {
+			break
+		}
+		if item.Pinned || item.Summary == "" || item.Content == item.Summary {
+			continue
+		}
+		oldTokens := item.Tokens
+		item.Content = item.Summary
+		item.Tokens = EstimateTokens(item.Content)
+		saved := oldTokens - item.Tokens
+		if saved > 0 {
+			s.usedTokens -= saved
+			result.Summarized++
+			result.TokensFreed += saved
+		}
+	}
+
+	// Phase 2: Deduplication — merge items with >70% word overlap
+	ids := make([]string, 0, len(s.items))
+	for id := range s.items {
+		ids = append(ids, id)
+	}
+	merged := make(map[string]bool)
+
+	for i := 0; i < len(ids); i++ {
+		if merged[ids[i]] {
+			continue
+		}
+		a := s.items[ids[i]]
+		if a == nil || a.Pinned {
+			continue
+		}
+		aWords := wordSet(a.Content)
+
+		for j := i + 1; j < len(ids); j++ {
+			if merged[ids[j]] {
+				continue
+			}
+			b := s.items[ids[j]]
+			if b == nil {
+				continue
+			}
+
+			if jaccardSimilarity(aWords, wordSet(b.Content)) > 0.7 {
+				// Keep higher-scoring item, merge tags
+				if s.scoreLocked(a) >= s.scoreLocked(b) {
+					a.Tags = mergeTags(a.Tags, b.Tags)
+					if a.Summary == "" && b.Summary != "" {
+						a.Summary = b.Summary
+					}
+					s.usedTokens -= b.Tokens
+					result.TokensFreed += b.Tokens
+					delete(s.items, ids[j])
+					merged[ids[j]] = true
+				} else {
+					b.Tags = mergeTags(b.Tags, a.Tags)
+					if b.Summary == "" && a.Summary != "" {
+						b.Summary = a.Summary
+					}
+					s.usedTokens -= a.Tokens
+					result.TokensFreed += a.Tokens
+					delete(s.items, ids[i])
+					merged[ids[i]] = true
+					break
+				}
+				result.Deduplicated++
+			}
+		}
+	}
+
+	// Phase 3: Evict lowest-scoring non-pinned items
+	if s.usedTokens > targetTokens {
+		evictable := make([]*Item, 0)
+		for _, item := range s.items {
+			if !item.Pinned {
+				evictable = append(evictable, item)
+			}
+		}
+		sort.Slice(evictable, func(i, j int) bool {
+			return s.scoreLocked(evictable[i]) < s.scoreLocked(evictable[j])
+		})
+
+		for _, item := range evictable {
+			if s.usedTokens <= targetTokens {
+				break
+			}
+			s.usedTokens -= item.Tokens
+			result.TokensFreed += item.Tokens
+			delete(s.items, item.ID)
+			result.Evicted++
+		}
+	}
+
+	// Collect items that could benefit from LLM summarization
+	for _, item := range s.items {
+		if item.Summary == "" && item.Tokens > 100 {
+			result.NeedsSummary = append(result.NeedsSummary, item)
+		}
+	}
+
+	result.TokensAfter = s.usedTokens
+	return result
+}
+
+// --- helpers ---
+
+func newID() string {
+	b := make([]byte, 8)
+	if _, err := rand.Read(b); err != nil {
+		panic("crypto/rand: " + err.Error())
+	}
+	return fmt.Sprintf("%x", b)
+}
+
+func wordSet(s string) map[string]struct{} {
+	words := strings.Fields(strings.ToLower(s))
+	set := make(map[string]struct{}, len(words))
+	for _, w := range words {
+		set[w] = struct{}{}
+	}
+	return set
+}
+
+func jaccardSimilarity(a, b map[string]struct{}) float64 {
+	if len(a) == 0 && len(b) == 0 {
+		return 0
+	}
+	intersection := 0
+	for w := range a {
+		if _, ok := b[w]; ok {
+			intersection++
+		}
+	}
+	union := len(a) + len(b) - intersection
+	if union == 0 {
+		return 0
+	}
+	return float64(intersection) / float64(union)
+}
+
+func hasAnyTag(item *Item, tags []string) bool {
+	tagSet := make(map[string]struct{}, len(item.Tags))
+	for _, t := range item.Tags {
+		tagSet[strings.ToLower(t)] = struct{}{}
+	}
+	for _, t := range tags {
+		if _, ok := tagSet[strings.ToLower(t)]; ok {
+			return true
+		}
+	}
+	return false
+}
+
+func mergeTags(a, b []string) []string {
+	seen := make(map[string]struct{})
+	var result []string
+	for _, t := range a {
+		lower := strings.ToLower(t)
+		if _, ok := seen[lower]; !ok {
+			seen[lower] = struct{}{}
+			result = append(result, t)
+		}
+	}
+	for _, t := range b {
+		lower := strings.ToLower(t)
+		if _, ok := seen[lower]; !ok {
+			seen[lower] = struct{}{}
+			result = append(result, t)
+		}
+	}
+	return result
+}
@@ -0,0 +1,200 @@
+package main
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestEstimateTokens(t *testing.T) {
+	tests := []struct {
+		input string
+		want  int
+	}{
+		{"", 0},
+		{"hi", 1},
+		{"hello world", 3},
+		{strings.Repeat("a", 100), 25},
+	}
+	for _, tt := range tests {
+		got := EstimateTokens(tt.input)
+		if got != tt.want {
+			t.Errorf("EstimateTokens(%q) = %d, want %d", tt.input, got, tt.want)
+		}
+	}
+}
+
+func TestStoreAddAndQuery(t *testing.T) {
+	s := NewStore(100000)
+
+	item := s.Add("cilium uses netkit on kernel 6.8", "cilium netkit needs 6.8", []string{"cilium", "networking"}, 8)
+	if item.ID == "" {
+		t.Fatal("expected non-empty ID")
+	}
+	if item.Tokens != EstimateTokens("cilium uses netkit on kernel 6.8") {
+		t.Errorf("tokens mismatch: got %d", item.Tokens)
+	}
+
+	// Query by text
+	results := s.Query("cilium kernel", nil, 10)
+	if len(results) != 1 {
+		t.Fatalf("expected 1 result, got %d", len(results))
+	}
+	if results[0].ID != item.ID {
+		t.Error("wrong item returned")
+	}
+
+	// Query by tag
+	results = s.Query("", []string{"networking"}, 10)
+	if len(results) != 1 {
+		t.Fatalf("expected 1 result by tag, got %d", len(results))
+	}
+
+	// Query with non-matching tag
+	results = s.Query("", []string{"nonexistent"}, 10)
+	if len(results) != 0 {
+		t.Fatalf("expected 0 results, got %d", len(results))
+	}
+}
+
+func TestStoreRemoveAndPin(t *testing.T) {
+	s := NewStore(100000)
+
+	item := s.Add("test content", "", nil, 5)
+	if !s.Pin(item.ID) {
+		t.Fatal("pin failed")
+	}
+
+	got, ok := s.Get(item.ID)
+	if !ok || !got.Pinned {
+		t.Fatal("item should be pinned")
+	}
+
+	if !s.Remove(item.ID) {
+		t.Fatal("remove failed")
+	}
+	_, ok = s.Get(item.ID)
+	if ok {
+		t.Fatal("item should be gone")
+	}
+}
+
+func TestCompaction(t *testing.T) {
+	s := NewStore(200) // small budget: ~200 tokens = ~800 chars
+
+	// Store items that exceed budget
+	s.Add("alpha bravo charlie delta echo foxtrot golf hotel", "short alpha", []string{"a"}, 3)
+	s.Add("india juliet kilo lima mike november oscar papa", "", []string{"b"}, 5)
+	s.Add("quebec romeo sierra tango uniform victor whiskey", "", []string{"c"}, 8)
+
+	_, used, _, _ := s.Status()
+	if used == 0 {
+		t.Fatal("expected non-zero usage")
+	}
+
+	result := s.Compact(0.5) // compact to 50%
+	if result.TokensAfter > result.TokensBefore {
+		t.Error("compaction should not increase tokens")
+	}
+	if result.TokensFreed == 0 && result.TokensBefore > 100 {
+		t.Error("expected some tokens freed")
+	}
+
+	// Item with summary should have been promoted
+	if result.Summarized == 0 {
+		t.Log("note: no summary promotions (may depend on budget math)")
+	}
+}
+
+func TestSummaryPromotion(t *testing.T) {
+	s := NewStore(100) // very tight budget
+
+	item := s.Add(
+		"this is a very long content string that takes many tokens to represent in the context window",
+		"long content, many tokens",
+		nil, 5,
+	)
+
+	result := s.Compact(0.3) // aggressive compaction
+	if result.Summarized == 0 {
+		t.Log("note: summary promotion did not trigger")
+	}
+
+	got, ok := s.Get(item.ID)
+	if ok && got.Content == "long content, many tokens" {
+		// Content was promoted to summary
+		if got.Tokens != EstimateTokens("long content, many tokens") {
+			t.Errorf("tokens should match promoted summary: got %d", got.Tokens)
+		}
+	}
+}
+
+func TestDeduplication(t *testing.T) {
+	s := NewStore(100)
+
+	s.Add("alpha bravo charlie delta echo foxtrot", "", []string{"a"}, 5)
+	s.Add("alpha bravo charlie delta echo golf", "", []string{"b"}, 5)
+
+	_, _, countBefore, _ := s.Status()
+	s.Compact(0.3)
+	_, _, countAfter, _ := s.Status()
+
+	if countAfter >= countBefore {
+		t.Log("note: dedup did not reduce count (similarity may be below threshold)")
+	}
+}
+
+func TestUpdateSummary(t *testing.T) {
+	s := NewStore(100000)
+
+	item := s.Add("full content here", "", nil, 5)
+	if item.Summary != "" {
+		t.Fatal("should have no summary initially")
+	}
+
+	if !s.UpdateSummary(item.ID, "compact version") {
+		t.Fatal("update failed")
+	}
+
+	got, _ := s.Get(item.ID)
+	if got.Summary != "compact version" {
+		t.Errorf("summary not updated: got %q", got.Summary)
+	}
+}
+
+func TestJaccardSimilarity(t *testing.T) {
+	a := wordSet("hello world foo bar")
+	b := wordSet("hello world foo baz")
+
+	sim := jaccardSimilarity(a, b)
+	if sim < 0.5 || sim > 0.7 {
+		t.Errorf("expected ~0.6 similarity, got %f", sim)
+	}
+
+	// Identical sets
+	if jaccardSimilarity(a, a) != 1.0 {
+		t.Error("identical sets should have similarity 1.0")
+	}
+
+	// Empty sets
+	if jaccardSimilarity(wordSet(""), wordSet("")) != 0 {
+		t.Error("empty sets should have similarity 0")
+	}
+}
+
+func TestAutoCompact(t *testing.T) {
+	s := NewStore(50) // very small: 50 tokens = ~200 chars
+	s.Configure(0, boolPtr(true), 0.5)
+
+	// Add items that should trigger auto-compaction
+	s.Add("first item with some content padding", "first", nil, 3)
+	s.Add("second item with more content padding", "second", nil, 3)
+	s.Add("third item triggering compaction now", "third", nil, 3)
+
+	_, used, _, usage := s.Status()
+	// Auto-compact should have kept usage reasonable
+	if usage > 1.0 {
+		t.Errorf("auto-compact should prevent exceeding budget: used=%d, usage=%.1f%%", used, usage*100)
+	}
+}
+
+func boolPtr(b bool) *bool { return &b }
@@ -0,0 +1,10 @@
+package main
+
+// EstimateTokens returns a rough token count for a string.
+// Uses ~4 characters per token heuristic.
+func EstimateTokens(s string) int {
+	if len(s) == 0 {
+		return 0
+	}
+	return (len(s) + 3) / 4
+}
@@ -0,0 +1,267 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"strings"
+
+	"github.com/mark3labs/mcp-go/mcp"
+	"github.com/mark3labs/mcp-go/server"
+)
+
+func registerTools(s *server.MCPServer, store *Store) {
+	s.AddTool(mcp.NewTool("store",
+		mcp.WithDescription("Store a context item for later retrieval. Offload information from working context. Provide a summary for efficient compaction when budget is tight."),
+		mcp.WithString("content", mcp.Required(), mcp.Description("The content to store")),
+		mcp.WithString("summary", mcp.Description("Compact summary of the content (used when budget is tight)")),
+		mcp.WithString("tags", mcp.Description("Comma-separated tags for categorization and retrieval")),
+		mcp.WithNumber("importance", mcp.Description("Importance 1-10, default 5. Higher = harder to evict")),
+	), handleStore(store))
+
+	s.AddTool(mcp.NewTool("query",
+		mcp.WithDescription("Retrieve stored context items by text search and/or tags. Returns most relevant items. Summaries returned when budget is tight."),
+		mcp.WithReadOnlyHintAnnotation(true),
+		mcp.WithDestructiveHintAnnotation(false),
+		mcp.WithString("query", mcp.Description("Text to search for in stored items")),
+		mcp.WithString("tags", mcp.Description("Comma-separated tags to filter by")),
+		mcp.WithNumber("limit", mcp.Description("Max items to return (default 10)")),
+	), handleQuery(store))
+
+	s.AddTool(mcp.NewTool("status",
+		mcp.WithDescription("Check context budget usage: total budget, tokens used, item count, usage percentage."),
+		mcp.WithReadOnlyHintAnnotation(true),
+		mcp.WithDestructiveHintAnnotation(false),
+	), handleStatus(store))
+
+	s.AddTool(mcp.NewTool("compact",
+		mcp.WithDescription("Trigger context compaction. Promotes summaries, deduplicates similar items, evicts lowest-scoring items. Returns suggestions for items needing LLM summarization."),
+		mcp.WithDestructiveHintAnnotation(true),
+		mcp.WithNumber("target_usage", mcp.Description("Target usage ratio 0.0-1.0 (default 0.7)")),
+	), handleCompact(store))
+
+	s.AddTool(mcp.NewTool("pin",
+		mcp.WithDescription("Pin a context item to prevent automatic eviction during compaction."),
+		mcp.WithDestructiveHintAnnotation(false),
+		mcp.WithIdempotentHintAnnotation(true),
+		mcp.WithString("id", mcp.Required(), mcp.Description("Item ID to pin")),
+	), handlePin(store))
+
+	s.AddTool(mcp.NewTool("forget",
+		mcp.WithDescription("Remove a context item from storage."),
+		mcp.WithDestructiveHintAnnotation(true),
+		mcp.WithString("id", mcp.Required(), mcp.Description("Item ID to remove")),
+	), handleForget(store))
+
+	s.AddTool(mcp.NewTool("configure",
+		mcp.WithDescription("Configure context budget and auto-compaction settings."),
+		mcp.WithDestructiveHintAnnotation(false),
+		mcp.WithIdempotentHintAnnotation(true),
+		mcp.WithNumber("token_budget", mcp.Description("Total token budget for storage")),
+		mcp.WithBoolean("auto_compact", mcp.Description("Enable/disable auto-compaction")),
+		mcp.WithNumber("auto_compact_threshold", mcp.Description("Auto-compact triggers at this usage ratio 0.0-1.0")),
+	), handleConfigure(store))
+
+	s.AddTool(mcp.NewTool("update",
+		mcp.WithDescription("Add or update the summary of an existing item. Use after compaction suggests items that need summarization."),
+		mcp.WithDestructiveHintAnnotation(false),
+		mcp.WithIdempotentHintAnnotation(true),
+		mcp.WithString("id", mcp.Required(), mcp.Description("Item ID to update")),
+		mcp.WithString("summary", mcp.Required(), mcp.Description("New summary for the item")),
+	), handleUpdate(store))
+}
+
+func handleStore(store *Store) server.ToolHandlerFunc {
+	return func(ctx context.Context, req mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+		content, err := req.RequireString("content")
+		if err != nil {
+			return mcp.NewToolResultError(err.Error()), nil
+		}
+
+		summary := req.GetString("summary", "")
+		tagsStr := req.GetString("tags", "")
+		importance := req.GetInt("importance", 5)
+
+		var tags []string
+		if tagsStr != "" {
+			for _, t := range strings.Split(tagsStr, ",") {
+				if t = strings.TrimSpace(t); t != "" {
+					tags = append(tags, t)
+				}
+			}
+		}
+
+		item := store.Add(content, summary, tags, importance)
+		budget, used, count, usage := store.Status()
+
+		return mcp.NewToolResultText(fmt.Sprintf(
+			"Stored [%s] (%d tokens)\nBudget: %d/%d tokens (%.0f%%), %d items",
+			item.ID, item.Tokens, used, budget, usage*100, count,
+		)), nil
+	}
+}
+
+func handleQuery(store *Store) server.ToolHandlerFunc {
+	return func(ctx context.Context, req mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+		query := req.GetString("query", "")
+		tagsStr := req.GetString("tags", "")
+		limit := req.GetInt("limit", 10)
+
+		var tags []string
+		if tagsStr != "" {
+			for _, t := range strings.Split(tagsStr, ",") {
+				if t = strings.TrimSpace(t); t != "" {
+					tags = append(tags, t)
+				}
+			}
+		}
+
+		items := store.Query(query, tags, limit)
+		if len(items) == 0 {
+			return mcp.NewToolResultText("No matching items found."), nil
+		}
+
+		tight := store.BudgetTight()
+		var sb strings.Builder
+		fmt.Fprintf(&sb, "Found %d items", len(items))
+		if tight {
+			sb.WriteString(" (budget tight, showing summaries where available)")
+		}
+		sb.WriteString(":\n\n")
+
+		for _, item := range items {
+			fmt.Fprintf(&sb, "[%s] importance:%d tokens:%d", item.ID, item.Importance, item.Tokens)
+			if item.Pinned {
+				sb.WriteString(" PINNED")
+			}
+			sb.WriteString("\n")
+			if len(item.Tags) > 0 {
+				fmt.Fprintf(&sb, "Tags: %s\n", strings.Join(item.Tags, ", "))
+			}
+			sb.WriteString("---\n")
+			if tight && item.Summary != "" {
+				sb.WriteString(item.Summary)
+			} else {
+				sb.WriteString(item.Content)
+			}
+			sb.WriteString("\n\n")
+		}
+
+		return mcp.NewToolResultText(sb.String()), nil
+	}
+}
+
+func handleStatus(store *Store) server.ToolHandlerFunc {
+	return func(ctx context.Context, req mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+		budget, used, count, usage := store.Status()
+		return mcp.NewToolResultText(fmt.Sprintf(
+			"Budget: %d/%d tokens (%.1f%%)\nItems: %d\nAuto-compact: %v (threshold: %.0f%%)",
+			used, budget, usage*100, count,
+			store.AutoCompact(), store.AutoCompactThreshold()*100,
+		)), nil
+	}
+}
+
+func handleCompact(store *Store) server.ToolHandlerFunc {
+	return func(ctx context.Context, req mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+		target := req.GetFloat("target_usage", 0.7)
+		if target <= 0 || target > 1.0 {
+			target = 0.7
+		}
+
+		result := store.Compact(target)
+
+		var sb strings.Builder
+		fmt.Fprintf(&sb, "Compaction complete:\n")
+		fmt.Fprintf(&sb, "- Evicted: %d items\n", result.Evicted)
+		fmt.Fprintf(&sb, "- Summary promoted: %d items\n", result.Summarized)
+		fmt.Fprintf(&sb, "- Deduplicated: %d pairs\n", result.Deduplicated)
+		fmt.Fprintf(&sb, "- Tokens freed: %d\n", result.TokensFreed)
+		fmt.Fprintf(&sb, "- Budget: %d → %d tokens\n", result.TokensBefore, result.TokensAfter)
+
+		if len(result.NeedsSummary) > 0 {
+			sb.WriteString("\nItems that would benefit from summarization:\n")
+			for _, item := range result.NeedsSummary {
+				preview := item.Content
+				if len(preview) > 80 {
+					preview = preview[:80] + "..."
+				}
+				fmt.Fprintf(&sb, "- [%s] %d tokens: %q\n", item.ID, item.Tokens, preview)
+			}
+			sb.WriteString("\nUse 'update' tool to add summaries to these items.")
+		}
+
+		return mcp.NewToolResultText(sb.String()), nil
+	}
+}
+
+func handlePin(store *Store) server.ToolHandlerFunc {
+	return func(ctx context.Context, req mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+		id, err := req.RequireString("id")
+		if err != nil {
+			return mcp.NewToolResultError(err.Error()), nil
+		}
+		if store.Pin(id) {
+			return mcp.NewToolResultText(fmt.Sprintf("Pinned [%s]", id)), nil
+		}
+		return mcp.NewToolResultError(fmt.Sprintf("Item [%s] not found", id)), nil
+	}
+}
+
+func handleForget(store *Store) server.ToolHandlerFunc {
+	return func(ctx context.Context, req mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+		id, err := req.RequireString("id")
+		if err != nil {
+			return mcp.NewToolResultError(err.Error()), nil
+		}
+		if store.Remove(id) {
+			_, used, count, usage := store.Status()
+			return mcp.NewToolResultText(fmt.Sprintf(
+				"Removed [%s]. Budget: %d tokens (%.0f%%), %d items",
+				id, used, usage*100, count,
+			)), nil
+		}
+		return mcp.NewToolResultError(fmt.Sprintf("Item [%s] not found", id)), nil
+	}
+}
+
+func handleConfigure(store *Store) server.ToolHandlerFunc {
+	return func(ctx context.Context, req mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+		budget := req.GetInt("token_budget", 0)
+		threshold := req.GetFloat("auto_compact_threshold", 0)
+
+		// Only set auto_compact if explicitly provided
+		var autoCompact *bool
+		if args := req.GetArguments(); args != nil {
+			if _, ok := args["auto_compact"]; ok {
+				v := req.GetBool("auto_compact", true)
+				autoCompact = &v
+			}
+		}
+
+		store.Configure(budget, autoCompact, threshold)
+
+		b, u, c, usg := store.Status()
+		return mcp.NewToolResultText(fmt.Sprintf(
+			"Configuration updated.\nBudget: %d/%d tokens (%.1f%%), %d items\nAuto-compact: %v (threshold: %.0f%%)",
+			u, b, usg*100, c,
+			store.AutoCompact(), store.AutoCompactThreshold()*100,
+		)), nil
+	}
+}
+
+func handleUpdate(store *Store) server.ToolHandlerFunc {
+	return func(ctx context.Context, req mcp.CallToolRequest) (*mcp.CallToolResult, error) {
+		id, err := req.RequireString("id")
+		if err != nil {
+			return mcp.NewToolResultError(err.Error()), nil
+		}
+		summary, err := req.RequireString("summary")
+		if err != nil {
+			return mcp.NewToolResultError(err.Error()), nil
+		}
+		if store.UpdateSummary(id, summary) {
+			return mcp.NewToolResultText(fmt.Sprintf("Updated summary for [%s]", id)), nil
+		}
+		return mcp.NewToolResultError(fmt.Sprintf("Item [%s] not found", id)), nil
+	}
+}