SuperLocalMemory 4.0: The Governed Memory Operating System for AI Agents
Governed memory with scoped component evidence and disclosed negative results
Memory · SuperLocalMemory
Store and recall project context through MCP or CLI. Keep the core memory data on infrastructure you control, choose operating modes explicitly, and inspect provenance. The research pages describe versioned protocols and assumptions.

Verify it yourself

Figures as reported by the project. Check them against the code and paper linked above.
Recipes
12 documented CLI workflows for storing context, recalling evidence, profiles, integrations and installation health. These are selected workflows rather than a claim of a complete command catalog.
Working proof
Carry project decisions and release requirements across agent sessions, then inspect the recalled evidence.
npm install -g superlocalmemoryslm setupslm doctor
On 1 October 2026, an isolated mode A CLI run stored and recalled one synthetic release-checklist fact. Calibration was uncalibrated and answer confidence was null. This is not a retrieval accuracy benchmark.
Install
Install
npm install -g superlocalmemoryWhat it does
Generation-fenced admission, policy registration, verifiable memory transactions, projection ownership, and hash-checkable completion manifests.
Dense, lexical, temporal, associative, and spreading-activation candidate producers fuse through one governed retrieval path.
Personal, named-profile, shared, and global scopes with role-aware access, provenance, export, and verified erasure.
Mode A local core, Mode B local model, and Mode C deliberate external provider usage with explicit network behavior.
Operate the same memory system from agent tools, terminal workflows, the dashboard, and framework adapters.
Feedback-driven learning, reusable retrieval paths, and context compression help agents retain useful context locally.
Use it for
Changelog
The latest releases on GitHub.
Three fixes for large stores, from reports on #306 and #304. No migration and no command to run.
Fixes for six reports filed after 4.2.11. No migration and no command to run.
Three fixes for things that looked done but were not. Upgrading runs one migration (M067, one trigger) and needs no command; a graph copy that already drifted is rebuilt by the daemon on its own within about ten minutes.
If you edited a memory (the dashboard's Edit button, the updatememory tool or slm update) and applied the correction, a question could no longer find it. The old text was retired and the new text was saved, but the new text was never added to the search index, so only exact…
If you run Mode B with Ollama, every saved memory used to cost up to eleven model calls, most of them wasted. A machine with a steady stream of memories kept Ollama busy all day, and recalls waited behind it. Nothing to run after upgrading.
Upgrading runs a few one-time repairs automatically the first time the daemon starts. Nothing to run. Every repair keeps an undo record, and slm doctor shows what each one did.
Research
Governed memory with scoped component evidence and disclosed negative results
Seven retrieval channels and zero LLM calls in the hot path; 70.4% on LoCoMo in zero-LLM mode
75% LoCoMo accuracy without cloud calls, from a four-channel retrieval architecture
Shared memory is an attack surface — Bayesian trust scoring that degrades a poisoned writer instead of trusting it