Skip to content
statefulmemory

Coding agents forget. statefulmemory remembers — Laya decides on-device.

Persistent memory for coding agents: per-project SQLite, hybrid search, code-anchored verify — plus a local System-1 (native MLX on Apple Silicon) that routes, decides, and judges conflicts in one forward pass. Self-host or use StatefulMemory Cloud.
Install Watch the film GitHub

brew install sanudatta11/tap/statefulmemory then smem install — wires your agent (MCP), git verify hooks, and the Laya System-1 sidecar. Skip Laya with –no-laya · same binary as sm

Watch the film

Section titled “Watch the film”

The full arc in two minutes — forget → remember → retrieve → verify → decide → connect → Laya.

What you get

Section titled “What you get”

14 install targets

Claude Code, Cursor, Windsurf, OpenCode, Codex, Gemini CLI, Amazon Q and more — MCP tools and skills wired in one command.

Inspectable SQLite

Per-project database under ~/.statefulmemory/ with FTS5. Read it, back it up, export it as a portable .mem archive — never a black box.

Hybrid, no API key

BM25 + dense retrieval fused with RRF, plus a local reranker on the self-host path. Search and context work fully offline.

On-device System-1

Native laya-mlx on Apple Silicon decides query tier, Decide, and conflict verdicts in one forward pass (~13 ms router P50 on an M3 Max). PyTorch elsewhere.

Chat history is not project memory

Section titled “Chat history is not project memory”

Your agent’s chat log dies with the window. Every new session, it re-learns the same decisions, re-trips the same pitfalls, and drifts on the same open questions. StatefulMemory keeps the durable record: what you decided, why, and which code it is pinned to — then hands the agent the right context the next time it starts.

Long context models still pay attention cost inside the window. Effective context compiles decisions, anchors, and facts into a fixed host window, so a 200k or 1M-token agent acts like a much larger working set.

How it works

Section titled “How it works”

Remember

smem obs save –type decision –title “use pgx not GORM” — every choice lands in per-project SQLite, optionally pinned to code.

Retrieve

smem obs context –query “which ORM should we use?” — hybrid BM25 + dense, token-budgeted, ranked for the task at hand.

Verify

smem obs save –anchor src/auth.rs::login … — git hooks re-check anchors; stale claims are withdrawn from context.

Decide

smem decide “SQLite or Postgres?” — conflicting memories get a typed verdict: in force, superseded, archived.

Connect

smem graph query auth –hops 2 — entity graph with PPR ranking surfaces related decisions you forgot existed.

Laya

smem install — a System-1 sidecar that answers router / decide / conflict questions locally, with silent fallback. Full story on the Laya page.

Laya System-1 — the on-device co-pilot

Section titled “Laya System-1 — the on-device co-pilot”

LAYA · SYSTEM-1

Thinks in one pass, not one round-trip.

The LLM is System-2: slow, token-hungry, remote. Laya is System-1: non-autoregressive, local, typed — it answers the discrete questions your memory layer asks (which retrieval tier, what should Decide say, does this conflict) in a single forward pass.

System-2

LLM round-trip

  • Tokens per call
  • Remote or paid
  • Latency on the network
System-1

Laya · one forward pass

  • Router ~30 ms / Decide ~100–200 ms on MLX
  • Loopback · no network
  • Open weights · Apache-2.0 · $0 self-hosted
query routerEasy skips dense · Hard expands + graph
deciderecommend a side · abstain when evidence is thin
conflict judgeSUPERSEDES verdicts for in-force memory
$ curl -s http://127.0.0.1:8765/health
{ "ok": true, "backend": "mlx", "loaded": [...], "error": null }

auto picks laya-mlx on Apple Silicon (macOS 14+, Python ≥ 3.11) and PyTorch elsewhere — same checkpoints, same typed schema. Weak calls soft-fail to heuristics or your agent CLI. Never a hard dependency. Read the Laya guide

Own your memory

Section titled “Own your memory”

Self-host

Daemon on your laptop (Unix socket) or team TCP+TLS. Data stays yours — inspect, backup, delete.

StatefulMemory Cloud

Same CLI and MCP surfaces, hosted storage and ops, when you would rather not run the daemon.

Portable

Export a .mem archive and import anywhere. Your memory follows you, not a vendor’s lock-in.

Start here

Section titled “Start here”
Why statefulmemoryHonest engineering comparison.InstallBuild from source in a minute.Wire into your agentMCP tools, git verify hooks.Laya System-1On-device typed decisions, native MLX.DecideResolve conflicting memories.ConfigOverlays, env overrides, backends.

For LLM crawlers: llms.txt · llms-full.txt · AGENTS.md (OpenCode / Codex / local) · CLAUDE.md.