OPEN SOURCE · AGPL V3 V3.7 System Map → Cache + Compression → V3.7 · LOCAL-FIRST AGENT MEMORY

Memory for
AI Reliability
Engineering

One local install for memory, retrieval, cache, compression, and agent coordination.
SuperLocalMemory records dated, attributable memory; retrieves through available semantic, keyword, temporal, associative, and graph channels; and exposes the same local system through MCP, CLI, hooks, and the dashboard.

3 public arXiv preprints · arXiv:2603.14588 · 2603.02240 · 2604.04514

Get Started Read the Preprints Explore the architecture

Need integration help? Contact SuperLocalMemory →

7
Operating layers
3
Operating modes A / B / C
5
Available recall producers
3
Public arXiv preprints

Registry downloads, not unique users: 4,565 npm downloads (15 Jun–14 Jul 2026) · 3,594 PyPI downloads (last 30 days, retrieved 16 Jul 2026).

superlocalmemory — zsh
V3.7 product architecture

One local memory system.
Every surface your agents use.

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.

LOCAL CONTROL PLANE

Canonical state stays local. Every external path is visible and explicitly enabled.

ALocal CoreNo cloud model required
BLocal OllamaOperator-managed local model
CExternal ProviderExplicit provider-backed enrichment
MEMORY RUNTIMEseven bounded stages
01

Admission

Identity, scope, idempotency, and raw evidence enter an inspectable operation.

02

Canonical store

SQLite, FTS, profile isolation, and operation receipts are the source of truth.

03

Enrichment

Entities, time, provenance, graph derivations, and lifecycle work advance in bounded stages.

04

Recall fusion

Healthy semantic, BM25, temporal, Hopfield, and graph candidates are fused with evidence.

05

Safe context

Budgets, redaction, provenance, and reference-only rendering protect agent injection surfaces.

06

Optimize

Exact cache, explicit invalidation, and safe compression retain operator control.

07

Operate

Diagnostics, policy, retention, export, backups, and health keep the runtime inspectable.

SQLite remains canonicalOptional systems never silently replace the local source of truth.
Inspect the full system →
01 · ingest and govern

Seven durable operating layers

  1. 01
    CaptureCLI, MCP, hooks, dashboard actions, and opt-in source adapters accept scoped content with source and time context.
  2. 02
    Durable recordRaw evidence, idempotency, profile, provenance, and dated facts enter the local operation state before enrichment.
  3. 02B
    Set memory boundariesProfiles isolate workspaces. Each memory is personal, shared with named profiles, or global; cross-profile recall remains default-deny until explicitly enabled.
  4. 03
    StructureFacts, entities, relationships, temporal signals, FTS, embeddings, and lifecycle state are built as configured dependencies are healthy.
  5. 04
    RetrieveSemantic, BM25, temporal, Hopfield, and spreading-activation channels contribute when available; fusion retains evidence provenance.
  6. 05
    Learn and governFeedback signals, ranking models, lifecycle, trust, retention, provenance, and audit records remain inspectable.
  7. 06
    OptimizeExact-result cache, optional semantic cache, safe compression, and opt-in aggressive prose compression reduce repeated context work.
  8. 07
    OperateDashboard, health, operations, entity explorer, skill evolution, settings, CLI, MCP profiles, and IDE hooks expose the system.
02 · choose an operating mode

The same memory contract, three execution paths.

MODE A · LOCAL COREDefault. Core memory operations use the configured local data root without a cloud model provider.
MODE B · LOCAL LLMAdds local Ollama enrichment and reranking while keeping the core memory path on the machine.
MODE C · EXTERNAL LLMAdds configured provider-backed enrichment. Queries or ingestion configured for that provider can leave the local system.

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.

Full package, not a feature list

What your team gets in the local runtime.

01 / MEMORY BOUNDARIES

Profiles and scoped memory

Profile-isolated workspaces plus personal, named-profile shared, and global scopes. Cross-profile recall is default-deny; Mesh is separate trusted-peer coordination.

02 / GRAPH

Knowledge graph and entities

Entity resolution, canonical entities, relations, graph exploration, and graph-informed recall evidence.

03 / BRAIN

Learning and ranking signals

Behavioral feedback, optional LightGBM model loading, score diagnostics, lifecycle state, and pattern views.

04 / EVOLUTION

Skill evolution

Skill lineage, guarded evolution workflows, evidence budgets, and operator-visible evolution state.

05 / OPTIMIZE

Cache and compression

Exact cache, tag invalidation, routed-result MCP cache, safe normalization, and opt-in lossy prose compression.

06 / MESH

Trusted local coordination

Authenticated peer messages, inbox, locks, offline queue, and optional mDNS discovery—not replicated distributed memory.

07 / AGENT SURFACES

MCP, CLI, hooks, and skills

Profile-selected MCP tools, structured CLI commands, and additive integrations for Claude Code, Codex, Cursor, Copilot, and Antigravity.

08 / OPERATIONS

Local operator console

Dashboard sections for memories, graph, brain, health, operations, entities, skills, Mesh, settings, and optimization.

09 / ADAPTERS

Opt-in source ingestion

Gmail, Calendar, and meeting transcript adapters are deliberately configured; they do not silently activate on install.

Published LoCoMo evidence carried into V3.7 arXiv:2603.14588 ↗
Mode A · Retrieval
74.8%
10 conversations · 1,276 questions · local retrieval with GPT-4.1-mini answer synthesis
PUBLISHED V3 PROTOCOL
Mode A · Raw
60.4%
10 conversations · 1,276 questions · local embeddings, retrieval, and zero-LLM answer construction
PUBLISHED V3 PROTOCOL
Mode C
87.7%
Conv-30 · 81 questions · cloud embeddings plus GPT-4.1-mini answer generation and judge
PUBLISHED V3 PROTOCOL

The V3 architecture under these published protocols is carried into V3.7. The numbers retain their original model, answer-construction, dataset, and sample scope; Mode B has no separate published LoCoMo run.

V3.7 — Control planes

Memory that can be inspected,
not merely invoked.

Recall quality, data handling, operating cost, and lifecycle behavior are separate controls—not promises hidden behind a single API call.

01 Retrieval

Evidence-led
recall.

  • Five available candidate channels
  • Fusion + optional reranking
  • Graph-informed evidence

SLM shows the evidence path behind recall. Healthy channels participate; unavailable dependencies degrade safely rather than inventing a result.

See the retrieval flow →
02 Lifecycle

Memory with
a lifecycle.

  • Active / warm / cold / archive states
  • Retention + provenance controls
  • Explicit operator actions

Lifecycle and retention are local, inspectable controls. Operators decide the policy and review their own backups, exports, and configured external systems.

See operating controls →
03 Cache

Reuse exact
responses.

  • Exact cache opt-in
  • Semantic cache experimental
  • Explicit MCP caching without a proxy

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 →
04 Compression

Content-specific
reduction.

  • Safe normalization by default
  • Optional reversible storage
  • Lossy prose mode is explicit

Safe mode preserves JSON and code and may produce no reduction. Aggressive prose compression is opt-in and lossy.

See compression controls →
Quick Start

One command.

The npm package creates a package-owned Python environment. Setup, client changes, model acquisition, and optional features remain explicit and inspectable.

terminal
# Install the published package (Node 18+ / npm 9+) $ npm install -g superlocalmemory
# Setup — choose your mode (A / B / C) $ slm setup
# Store your first memory $ slm remember "Alice works at Google as Staff Engineer" --json ↳ state: queryable · operation: <opaque-id> · enrichment pending
# Recall it later $ slm recall "What does Alice do?" ↳ Staff Engineer at Google · relevance_score: 0.82 · memory_confidence: 0.97 score_contract_version: 2 · calibration_status: uncalibrated · answer_confidence: null
✓ V3 engine active · Mode A · local data root verified · cache: opt-in
Install Now Choose a Mode

AGPL v3 · Local-first core · Optional network paths documented

Mathematical Foundation

Geometry, not guesswork.

Three public arXiv preprints document versioned experiments. They are not venue-reviewed, and current runtime claims require release-linked proof.

01 · Retrieval metric
dFR(p, q) = arccos( Σ √(pi · qi) )

Fisher-Rao Distance

Dense candidate generation uses cosine similarity. Fisher-derived terms can inform later scoring when their state is available.

arXiv:2603.14588 →
02 · Lifecycle model
∇γ ∂tγ = 0  ·  Expp(v)

Riemannian Lifecycle

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 →
03 · Compression theory
H(X|Y) ≤ H(X)  ·  I(X;Y) ≥ 0

Information-Theoretic Compression

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 →
2603.14588 — V3 SLM: Information-Geometric Agent Memory 2603.02240 — Privacy-Preserving Multi-Agent Memory 2604.04514 — The Living Brain
Integrations

Documented client configurations.

Named templates are configuration surfaces, not proof of a complete integration. Run slm connect --list for the published package surface and consult the V3.7 release matrix before calling any client verified.

Claude Code Cursor MCP Protocol ChatGPT Desktop Perplexity VS Code Copilot Windsurf Continue.dev Zed GitHub Actions Docker n8n Goose OpenClaw Codex nanobot Shell scripts

Data-returning CLI commands document --json structured output where supported.
Consumers should parse versioned fields, not display text.

Common Questions

Frequently Asked Questions