01
Admission
Identity, scope, idempotency, and raw evidence enter an inspectable operation.
Memory for AI Reliability Engineering. SuperLocalMemory is a governed, local-first memory system for AI agents and teams — memory runs on infrastructure you control.
It records dated, attributable memory; retrieves through available semantic, keyword, temporal, associative, and graph channels; and provides scoped memory, role-based access, and GDPR / EU AI Act governance controls. V4.0 adds bounded loops, framework adapters, and a multi-agent memory view through MCP, CLI, hooks, and the dashboard.
4 public arXiv preprints · arXiv:2608.08253 · 2603.14588 · 2603.02240 · 2604.04514
Need integration help? Contact SuperLocalMemory →
V4.0 adds three primitives for AI Reliability Engineering: gate-verified iteration, a shared memory backend for documented agent frameworks, and an operational view of who wrote what.
Agent iteration that terminates when an independent gate passes — not when the agent self-reports. Budget bounds (iteration cap, token budget, wall-clock) are enforced by the engine. Every lap writes a durable ledger entry to your SLM data root.
slm loop CLI · /slm-loop command · MCP tools (slm_loop_run, slm_loop_history, slm_loop_show)LangGraph, Semantic Kernel, Microsoft Agent Framework, LangChain, LlamaIndex, CrewAI, AutoGen, Google ADK, and OpenAI Agents. Each wires the framework's native memory interface to your local SLM data root. Adapters make no outbound connections; optional network features are separate opt-ins.
See the adapters table →A dedicated dashboard page shows per-agent memory attribution, bounded-loop run history with per-lap verdicts, and the registered agent list. It is available at your SLM daemon address in V4.0.
Admin / member / viewer roles per workspace, sign-in for shared workspaces, and per-workspace tenant isolation. GDPR access, erasure, and portability with a hash-chained audit trail; optional PII redaction on save (email, phone, national ID, payment card, IP); and a per-mode EU AI Act self-assessment.
See governance & EU AI Act controls →Read the release history at superlocalmemory.com/changelog.
SLM is not a vector-store wrapper. It is a durable memory pipeline, a retrieval pipeline, and an operator surface for the agents, IDEs, and local services around it.
Canonical state stays local. Every external path is visible and explicitly enabled.
Identity, scope, idempotency, and raw evidence enter an inspectable operation.
SQLite, FTS, profile isolation, and operation receipts are the source of truth.
Entities, time, provenance, graph derivations, and lifecycle work advance in bounded stages.
Healthy semantic, BM25, temporal, Hopfield, and graph candidates are fused with evidence.
Budgets, redaction, provenance, and reference-only rendering protect agent injection surfaces.
Exact cache, explicit invalidation, and safe compression retain operator control.
Diagnostics, policy, retention, export, backups, and health keep the runtime inspectable.
Storage truth: SQLite with WAL, FTS, and derived local indexes is the canonical data path. The Scale Engine can stage and verify CozoDB graph and LanceDB vector projections before explicit promotion. Optional connectors, backups, model downloads, and Mesh peers have separate network behavior.
EU AI Act mode check: a built-in self-assessment maps each mode to the framework — Modes A and B keep memory processing local with no generative AI; Mode C uses an external provider. It is a technical-control map, not a certification.
Profile-isolated workspaces plus personal, named-profile shared, and global scopes. Cross-profile recall is default-deny; Mesh is separate trusted-peer coordination.
Entity resolution, canonical entities, relations, graph exploration, and graph-informed recall evidence.
Behavioral feedback, optional LightGBM model loading, score diagnostics, lifecycle state, and pattern views.
Skill lineage, guarded evolution workflows, evidence budgets, and operator-visible evolution state.
Exact cache, tag invalidation, routed-result MCP cache, safe normalization, and opt-in lossy prose compression.
Authenticated peer messages, inbox, locks, offline queue, and optional mDNS discovery—not replicated distributed memory.
Profile-selected MCP tools, structured CLI commands, and additive integrations for Claude Code, Codex, Cursor, Copilot, and Antigravity.
Dashboard sections for memories, graph, brain, health, operations, entities, skills, Mesh, settings, and optimization.
Gmail, Calendar, and meeting transcript adapters are deliberately configured; they do not silently activate on install.
The V3 architecture under these published protocols is carried into V4.0. The numbers retain their original model, answer-construction, dataset, and sample scope; Mode B has no separate published LoCoMo run.
Recall quality, data handling, operating cost, and lifecycle behavior are separate controls—not promises hidden behind a single API call.
SLM shows the evidence path behind recall. Healthy channels participate; unavailable dependencies degrade safely rather than inventing a result.
See the retrieval flow →Lifecycle and retention are local, inspectable controls. Operators decide the policy and review their own backups, exports, and configured external systems.
See operating controls →Exact caching is available when Optimize or proxy caching is enabled. MCP and skill surfaces cache only content explicitly routed through SLM.
See cache controls →Safe mode preserves JSON and code and may produce no reduction. Aggressive prose compression is opt-in and lossy.
See compression controls →The npm package creates a package-owned Python environment. Setup, client changes, model acquisition, and optional features remain explicit and inspectable.
AGPL v3 · Local-first core · Optional network paths documented
Four public arXiv preprints document versioned experiments. They are not venue-reviewed, and current runtime claims require release-linked proof.
Dense candidate generation uses cosine similarity. Fisher-derived terms can inform later scoring when their state is available.
arXiv:2603.14588 →Lifecycle state changes combine explicit policy, observed use, decay, and optional research-informed dynamics. Runtime behavior is defined by the released code and tests.
arXiv:2604.04514 →Information theory describes limits, not a shipped compression ratio. Safe mode preserves JSON and code and may produce no reduction; reversible storage is verified separately.
Product evidence boundary →Named templates are configuration surfaces, not proof of a complete integration. Run slm connect --list for the published package surface and consult the V4.0 release matrix before calling any client verified.
Nine framework-specific adapters wire each framework's native memory interface to your local SLM data root. AI Reliability Engineering requires a single auditable store across your agent system, regardless of which framework each agent runs in. Every adapter is a thin binding — no new storage layer, no separate service. The same memories are visible through the CLI, MCP tools, and the dashboard.
| Framework | Package | What it wires |
|---|---|---|
| LangGraph | langgraph-superlocalmemory | BaseStore long-term memory |
| Semantic Kernel | semantic-kernel-superlocalmemory | VectorStore / RecordStore |
| Microsoft Agent Framework | agent-framework-superlocalmemory | Memory + conversation history providers |
| LangChain | langchain-superlocalmemory | BaseChatMessageHistory + retriever |
| LlamaIndex | llamaindex-superlocalmemory | Chat store |
| CrewAI | crewai-superlocalmemory | ExternalMemory storage backend |
| AutoGen | autogen-superlocalmemory | Memory provider |
| Google ADK | google-adk-superlocalmemory | Memory provider |
| OpenAI Agents | openai-agents-superlocalmemory | Memory provider |
Each adapter stores data in your local SLM data root. Optional SLM providers, backup, and connectors have separate, opt-in network behavior — the adapters themselves make no outbound connections.
Data-returning CLI commands document --json structured output where supported.
Consumers should parse versioned fields, not display text.