Why statefulmemory

StatefulMemory is persistent memory infrastructure for coding agents: CLI + MCP → gRPC daemon → per-project SQLite (FTS5 + optional hybrid). You can self-host (laptop or team TCP) or use StatefulMemory Cloud (managed SaaS).

This page is the place for an honest engineering comparison. Product UI and CLI help do not roast peers by name.

What you get

• Self-host: inspectable per-project SQLite under ~/.statefulmemory/
• Cloud SaaS: managed hosting for teams that do not want to run the daemon
• Hybrid BM25 + dense retrieval on the self-host path without a search API key
• Fourteen statefulmemory install targets plus seven MCP tools (statefulmemory / smem / sm are the same binary)
• Code anchors + git verify (stale withdrawal from context)
• Optional LLM steps via your agent CLI
• Effective context (smem context compile): pack decisions / anchors / facts into a fixed host window (8k…1m) so a 200k or 1M agent acts like a much larger working set — durable store stays outside the KV cache
• Repo ingest (smem ingest repo): path-anchored observations from git-tracked files
• Opt-in entity graph with BFS or HippoRAG-style PPR ranking (graph.ranker = ppr)
• Wave 3 retrieve: adaptive Easy/Normal/Hard router, local MiniLM-style CE rerank (default, LLM-free), result LRU cache, index-time fact key expand
• Laya System-1 (opt-in): on-device typed decisions for router / decide / conflict — native MLX on Apple Silicon, PyTorch elsewhere. Not a reranker; soft-fails to heuristic (router) or agent CLI (decide/conflict). See tools/laya-sidecar/ and the Laya guide.

Effective context (200k / 1M ≈ 100M with memory)

Long-context models still pay attention cost inside the window. StatefulMemory keeps the corpus in SQLite and compiles a slot-budgeted brief (≈35% decisions, ≈35% anchored code, ≈30% related) for whatever window the host exposes. Retrieval borrows LongMemEval indexing tips (fact-augmented query merge, heuristic time pruning) and optional Personalized PageRank over the entity graph for single-step multi-hop — without claiming Magic-style 100M trained LTM weights.

Unmatched-harness caveat: this is a systems claim (more useful tokens per window), not a published LoCoMo/LongMemEval scorecard win.

Laya System-1: decisions, not text

The optional Laya sidecar answers the discrete questions the memory layer asks — which retrieval tier a query needs, what Decide should recommend, whether two memories conflict — in one non-autoregressive forward pass. On Apple Silicon it runs native laya-mlx (published M3 Max P50 13.4 ms for the 421M checkpoint, no PyTorch); elsewhere it uses the PyTorch laya package. Same checkpoints, same typed schema. Weak calls soft-fail to heuristics or your agent CLI, so search and context keep working with Laya off.

Numbers are from the laya-mlx project’s published benchmarks and are not a statefulmemory-harness score. See Laya System-1.

Comparison

Dimension	StatefulMemory	Hosted memory APIs	Vector DB alone
Fully local / self-host save / search / context	✓	Often no	If you self-host
Managed Cloud SaaS	✓ (offering)	Yes	N/A
Coding-agent install (MCP / skills)	✓ (14 targets)	Rare	No
Inspectable per-project SQLite (self-host)	✓	No	N/A
Hybrid BM25 + dense on self-host	✓	Varies	Vectors only
Code anchors + git verify	✓	No	No
Python Memory SDK	Not yet	Often yes	Client libs

Peer benchmarks (caveats)

Published LoCoMo-style LLM-judge numbers for other systems (for example Mem0 ~66.9%, Engram-family bands roughly 68–80%) come from unmatched harnesses — different judges, category filters, and retrieval settings. StatefulMemory does not claim to beat those numbers until a disclosed stratified scorecard exists.

How we run and interpret evals: LoCoMo eval.

What we do not claim

• No Python / TypeScript Memory SDK today
• No LangGraph / CrewAI / LlamaIndex as shipped integrations
• No Windows support for the CLI v1
• No “beats peer X” until stratified LoCoMo numbers are published

Next: Architecture · Integrations · Self-hosting and Cloud.