feat: add persistent context memory store with write-time dedup

[Siddhant-K-code]

Adds a persistent context memory store that accumulates knowledge across agent sessions. Memories are deduplicated on write, ranked by relevance + recency on recall, and compressed over time through hierarchical decay.

What it does

Store - Write context with automatic dedup. If a semantically similar memory already exists (cosine distance < threshold), the existing one is updated instead of creating a duplicate.

Recall - Retrieve memories ranked by (1-w)*similarity + w*recency. Supports tag filtering and token budgets.

Forget - Remove memories by ID, tag, or age.

Decay - Background worker compresses aging memories:

• Full text -> summary (extractive, ~20% of original) after 24h
• Summary -> keywords after 7 days
• Keywords -> evicted after 30 days
• Accessing a memory resets its decay clock

New files

File	Purpose
pkg/memory/store.go	Interface, types, config
pkg/memory/sqlite.go	SQLite backend with write-time dedup
pkg/memory/decay.go	Hierarchical decay worker
pkg/memory/helpers.go	ID generation, embedding encode/decode, token estimation
pkg/memory/memory_test.go	11 tests covering store, recall, dedup, forget, decay, tags, token budget
cmd/memory.go	CLI: distill memory store/recall/forget/stats
cmd/api_memory.go	HTTP: POST /v1/memory/store, /recall, /forget, GET /stats
cmd/mcp_memory.go	MCP tools: store_memory, recall_memory, forget_memory, memory_stats

Dependencies

• modernc.org/sqlite - Pure Go SQLite (no CGO required)

Usage

# CLI
distill memory store --text "Auth uses JWT with RS256" --tags auth --source docs
distill memory recall --query "How does auth work?" --max-results 5
distill memory stats

# API
curl -X POST localhost:8080/v1/memory/store -d '{
  "session_id": "session-1",
  "entries": [{"text": "Auth uses JWT", "tags": ["auth"]}]
}'

# MCP (available as tools in Claude Desktop, etc.)
# store_memory, recall_memory, forget_memory, memory_stats

Closes #29

[Siddhant-K-code]

Code Review

CI is green (build + lint). 11 tests pass. Here's a deep review.

What's good

• Clean interface design - Store interface in store.go is minimal and backend-agnostic. Adding Redis or Postgres backends later is straightforward.
• Write-time dedup - Deduplicating on store (not recall) is the right call. It keeps the DB small and makes recall fast.
• Decay model - The three-level decay (full -> summary -> keywords -> evict) with access-count-based clock reset is well thought out.
• Pure Go SQLite - modernc.org/sqlite avoids CGO, making cross-compilation and Docker builds simpler.
• Consistent patterns - Follows the existing codebase conventions (error types, config structs, CLI flag patterns).
• Test coverage - 11 tests covering the main paths: store, dedup, recall, tags, token budget, forget, decay, edge cases.

Issues to address (in follow-ups)

P1 - Bug risk:

• touchMemories goroutine uses the request context which may be cancelled after response. Should use context.Background().

P2 - Correctness:

• MaxMemories config field is defined but never enforced in Store(). Either implement the eviction or remove the field.
• Tag filtering uses LIKE '%"tag"%' on JSON arrays - substring matching can produce false positives (tag "au" matches "auth").

P3 - Performance (acceptable now, track for later):

• findDuplicate does O(N) full table scan on every store. Fine for <1k memories, needs an index for larger stores.
• Memory store is always created on distill api startup even when unused. Should be opt-in.

P4 - Code quality:

• embeddingProvider interface in api_memory.go duplicates retriever.EmbeddingProvider.
• extractSummary in decay.go is position-based only. The existing pkg/compress/extractive.go has a better scorer that could be reused.

Summary

Solid first implementation. The core store/recall/forget/decay loop works correctly and the integration across CLI, API, and MCP is complete. The P1 bug and P2 items should be addressed before merging to main; P3/P4 can be tracked as follow-up issues.

[Siddhant-K-code]

Updated Review - Post Refactor

CI passing (build + lint), all 11 tests pass, 0 lint issues. Previous P1-P4 findings addressed.

Verdict: Approve with minor suggestions

No blocking issues. Architecture is sound, single-connection SQLite pattern is correctly handled, test coverage is good.

Remaining Findings

Priority	File	Finding
P2	sqlite.go:143-157	findDuplicate does a full table scan of all embeddings. O(n) per insert. Fine for < 10K entries, add a TODO for scaling.
P3	sqlite.go:296-310	Stats has sequential defer rows.Close() - works but fragile with single-conn. Use scan-then-process for consistency.
P3	decay.go:120-131	extractSummary allocates a new compressor per call. Make it a package-level var.
P3	mcp.go:159	Memory DB path hardcoded in MCP. API server reads from viper. Add --memory-db flag.
Nit	sqlite.go:39-42	WAL mode has no benefit with MaxOpenConns(1). Remove or comment.
Nit	sqlite.go:82-84	PRAGMA foreign_keys should be in NewSQLiteStore alongside other PRAGMAs, not in migrate().
Nit	decay.go:148-161	isStopWord rebuilds map on every call. Move to package-level var.
Nit	memory.go:260	Trailing blank line.

What's Good

• Junction table for tags - correct approach
• Scan-then-process in Recall - avoids single-connection deadlock
• --memory opt-in flag - clean separation
• retriever.EmbeddingProvider reuse - no duplicate interface
• Extractive compressor reuse in decay
• 11 tests covering all key paths
• Clean type separation (StoreEntry / Entry / RecalledMemory)