v1.4.4

What's New in v1.4.4

Memory Synthesis

New memstack.intelligence.synthesis module — the LLM now synthesises auto-captured exchanges into concise memories instead of storing raw conversation turns. This reduces noise and keeps the vault focused on what matters.

Background Consolidation

New memstack.intelligence.consolidation module — a background task periodically reviews stored memories, merging duplicates and pruning stale entries. Prevents vault drift over time. Includes fixes for event loop blocking and split data loss.

Version & Docs

Version bumped to 1.4.4 across pyproject.toml and __init__.py. README rewritten in condensed reference format. Dev/reference files excluded from release tracking.

Full Changelog

See CHANGELOG.md for complete details.

v1.4.3

What's New in v1.4.3

Merge Pipeline

LLM can now decide to merge incoming content into an existing memory instead of only add, update, or ignore. Merged memories append content with a \n\n separator, preserving the original ID, importance, and created timestamp. New VaultStore.merge() method powers the append logic.

Similarity Thresholds

Lowered similarity_add_threshold (0.3 → 0.25) and similarity_ignore_threshold (0.92 → 0.85) to reduce false deduplication. High-similarity memories now go to the LLM for a decision instead of being auto-ignored by default.

Auto-Ignore Toggle

New similarity_ignore_enabled config field (default: false). When true, restores the old auto-ignore behavior for scores at or above the ignore threshold. When false (default), all scores at or above the add threshold go to the LLM for a decision.

Auto-Capture Improvements

OpenClaw bridge now captures all messages instead of only the last 4. Tool role messages are included alongside user and assistant. Structured content blocks (tool_use, tool_result, text arrays) are serialized as readable text instead of [object Object]. System messages are correctly excluded.

Release Cleanup

Updated .gitignore to exclude internal dev and continuity files (CONTEXT.md, STATUS.md, progress.md, reference/, .claude/).

Full Changelog

See CHANGELOG.md for complete details.

v1.4.2

What's New in v1.4.2

Caching Layer

Ollama client singleton, LRU embedding cache, FTS rebuild interval, TTL search result cache, and LRU vault read/list cache — 4 new config fields (embedding_cache_size, fts_rebuild_interval, search_cache_ttl, vault_cache_size).

Async Handlers

All REST handlers and MCP tools converted to async def with run_in_executor for blocking sync calls. Importance updates deferred via asyncio.create_task() — responses return immediately.

System Prompt Fix

Sections 0, 1, and 6 rewritten with factual tone. Section 6 now states memories are injected automatically instead of instructing models to call memory_inject before every response.

OpenClaw Bridge Cache

TTL cache (30s) on before_prompt_build hook keyed by agent ID + first 500 chars of prompt — identical prompts within the window skip the MemStack HTTP call.

Bug Fixes

• Shared memory importance updates now skip vault.update_importance for reserved "shared" agent_id
• Vault cache thread safety via threading.Lock

Full Changelog

See CHANGELOG.md for complete details.

v1.4.1

What's New in v1.4.1

MCP Standalone Server

MCP now runs as a separate process on port 7778 (MEMSTACK_MCP_PORT) instead of being mounted inside the REST API. The CLI manages both processes automatically.

System Prompt Endpoint

GET /agents/{agent_id}/system-prompt returns the behavioral mandate and 7-tool reference. A matching memory_get_system_prompt MCP tool provides identical output.

OpenClaw Bridge

TypeScript plugin with auto-recall, auto-capture, and native memory blocking hooks. Includes a step-by-step connection guide (docs/connect-openclaw.md).

Bug Fixes

• MCP subprocess now receives --host and --port flags from CLI daemon launcher
• Undefined agentId guards in bridge hooks prevent "undefined" namespace pollution

Full Changelog

See CHANGELOG.md for complete details.

v1.4.0

What's New in v1.4.0

Shared Mode

Private-first write routing with shared copies. When MEMSTACK_SHARED_MODE=true, private writes are also copied to a shared directory accessible to all agents. Reserved shared agent_id is rejected in validation. File-level threading locks for concurrent write safety.

Inject Endpoint

GET /agents/{agent_id}/inject?q=&limit= and GET /shared/inject?q=&limit= — auto-injection that searches for relevant memories and returns top-N results above injection_min_score. MCP memory_inject tool with the same logic. Merges agent and shared namespace results when shared mode is enabled.

Bug Fixes

• Slug truncation to 80 chars prevents OSError on long content
• MCP path doubling fix (/mcp/mcp → /mcp)

Evals

v1.4 retrieval evaluation — P@1: 0.870, MRR@10: 0.924, R@5: 1.0

New Config

• MEMSTACK_INJECTION_MIN_SCORE (default 0.3)
• MEMSTACK_INJECTION_TOP_N (default 5)
• MEMSTACK_SHARED_MODE (default false)

---

Full Changelog: v1.3.0...v1.4.0

v1.3.0

MCP server, file watcher, search fixes, Windows daemon

What's new

• FastMCP server — 5 tools (memory_write, memory_search, memory_read, memory_delete, memory_list) mounted as sub-app at /mcp within the FastAPI process. Input validation for path traversal and invalid identifiers. Search results capped at 100.
• File watcher — real-time vault change detection via watchfiles. Incremental scan_vault() on startup with mtime-based scanning. Self-write suppression so vault writes don't trigger re-indexing loops.
• Smart write pipeline — LLM consultation for ambiguous similarity zones via local Ollama.
• Importance decay & hit increment — lazy importance updates at retrieval time.
• system_prompt_kit.md — agent-facing system prompt with behavioral rules and MCP tool reference.
• 4 new config fields — WATCHER_ENABLED, WATCHER_DEBOUNCE_MS, MCP_ENABLED, MCP_PATH

Evaluation results

Metric	v1.2	v1.3	Change
P@1	0.6957	0.6522	-6%
R@5	1.0	1.0	=
MRR@10	0.8152	0.7877	-3%
Latency@10	18561ms	18843ms	+2%

Precision and MRR dipped slightly due to MCP endpoint routing changes and watcher startup overhead. Latency essentially unchanged.

Bug fixes

• CLI crash when state file has last_indexed but no pid key
• SearchIndex.add() silently returned None on failure — now returns bool
• scan_vault() skipped failed files permanently — now retries on next startup
• Windows daemon PID liveness checks used os.kill (broken with PID reuse) — replaced with Win32 API via ctypes
• Windows daemon missing CREATE_NO_WINDOW flag and stdin=DEVNULL
• .env.example merge conflict residue

Full Changelog: v1.2.0...v1.3.0

v1.2.0

Smart write, importance scoring, evals

What's new

• Smart write pipeline — every memory write passes through embed → find_similar → threshold check → LLM consultation → execute decision. Three outcomes: add (new memory), update (replace similar memory preserving importance), ignore (duplicate).
• Importance scoring — memories have importance scores that decay over time (half-life default 7 days) and get bumped on retrieval (default +0.05 per hit). Decay is lazy, computed only at retrieval time.
• LLM consultation — when similarity is ambiguous (between add and ignore thresholds), a local Ollama model decides whether to add, update, or ignore. Falls back to "add" if Ollama is unreachable.
• 6 new config fields — SIMILARITY_ADD_THRESHOLD, SIMILARITY_IGNORE_THRESHOLD, IMPORTANCE_DECAY_HALFLIFE, IMPORTANCE_HIT_INCREMENT, LLM_MODEL, LLM_HOST
• MemoryWriteResponse — write endpoint returns {decision, id, similarity_score} with status 201 (added) or 200 (updated/ignored)

Evaluation results

Metric	v1.1	v1.2	Change
P@1	0.3478	0.6957	+100%
R@5	1.0	1.0	=
MRR@10	0.5841	0.8152	+39%
Latency@10	5324ms	18561ms	+249%

Precision and MRR improved significantly with smart write deduplication and importance-weighted reranking. Latency increased due to importance update on retrieval.

Bug fixes

• Pipeline data-loss (index delete after vault write)
• LLM host not configurable
• Empty importance_updated crash
• Inconsistent agent_id validation
• Health version mismatch
• Data loss when update produces same slug as original

Full Changelog: v1.1.0...v1.2.0

v1.1.0

What's New in v1.1.0

Search & Embeddings

• Hybrid search — vector + full-text via LanceDB with importance-weighted RRF reranking
• Semantic chunking — paragraph-first split, sentence split, hard token ceiling with tiktoken
• Embedding providers — Ollama and fastembed with automatic fallback; graceful degradation when unavailable
• Search endpoint — GET /agents/{agent_id}/memories/search?q=...&limit=...

CLI

• Daemon mode — memstack start --daemon with cross-platform process management
• Stop command — memstack stop with proper Windows (taskkill) and Unix (SIGTERM) support

Configuration

• 7 new settings: embedding_provider, embedding_model, embedding_autofallback, chunk_max_tokens, chunk_overlap_tokens, rrf_k, importance_rerank_weight

Evaluation Framework

• Golden-set evaluation with precision/recall/MRR metrics
• Parameter sweep runner for chunking config optimization

Bug Fixes

• Daemon timeout 5s → 30s (first-start crash)
• Windows onnxruntime DLL failure: graceful fallback instead of crash
• Windows memstack stop uses taskkill directly
• Test isolation from .env file

Full Changelog: v1.0.0...v1.1.0

v1.0.0

MemStack is a local-first memory server for AI agents. Memories are stored as Markdown files with YAML frontmatter in a vault directory you own. No cloud, no database, no external services. You can read and edit every memory with any text editor or Obsidian. The vault is ground truth.

What's included (v1.0.0)

• Vault storage — per-agent directory namespacing, atomic writes, path traversal protection
• REST API — 5 endpoints on FastAPI/Uvicorn: health check, memory CRUD under /agents/{agent_id}/memories
• CLI — memstack start (foreground + daemon) and memstack stop with cross-platform process management
• Configuration — 12 environment variables via Pydantic Settings, .env file support
• Slug-based IDs — deterministic from content + agent + date
• 102 tests, 95.5% coverage