MemStore v1.0

Fully model-free durable memory substrate for autonomous agents.

MemStore gives AI agents persistent, queryable memory with built-in temporal truth, contradiction detection, versioned snapshots with branching and three-way merge, multi-agent sync, trust scoring, online-learned retrieval weights, and budget-aware evidence compression — over exactly 14 MCP tools. No models, no downloads, eager startup, fully offline.

Architecture & database mapping: docs/architecture.md and docs/architecture.html.

Install (production)

Recommended for users: the release zip + install.py (no signatures needed, no SmartScreen):

Expand-Archive .\Downloads\memstore_v*_*.zip -DestinationPath C:\MemStore
python C:\MemStore\memstore_v2\install.py

The installer is interactive by default (numbered choices for install type, tenant data dir, LMDB map size, license key) and fully scriptable (python install.py --type full --tenant-dir E:\data\tenants --yes). It writes a valid config.yaml, creates Start-menu shortcuts, writes uninstall.cmd (keep-or-delete data prompt) — and it is non-intrusive: no IDE/CLI configs are touched; it generates MCP-CONFIGURATION.md with copy-paste connection snippets for every AI client detected so you wire them yourself (or later via clients-register --dry-run, inspect, apply explicitly).

Developer path: python install.py --extras dev (repo venv + smoke test) or pip install -e ".[dev]" + python -m memstore --transport stdio.

Health check: python -m memstore doctor Release zip: python -m memstore build

Updating an existing install

Run the NEW zip's installer against the existing install — no manual extract-over needed:

python C:\Downloads\memstore_v2\install.py --update C:\Downloads\memstore_v1.0.1.zip --dir C:\MemStore\memstore_v2

Conflict-safe by design: files you modified are never overwritten (the new version lands next to yours as name.new-<version> with a merge report), config.yaml, the venv, logs and all tenant data are untouched, every replaced file is backed up to backups/pre-update-<timestamp>.zip first, and re-running the same update is a no-op.

Two folders: program vs. memories

The install folder holds only the program. Your memories live in the tenant data dir you choose during setup (Browse… button in the wizard, default %LOCALAPPDATA%\MemStore\data\tenants):

• tenant_<id>.db — one LMDB database per tenant, created lazily on first store; tenant_<id>_<branch>.db for branched histories.
• backups/, weights/, feedback/ — snapshots, EMA weights, feedback log.

Every client connected by the installer gets its memstore MCP entry pinned to that tenant data dir via env = { MEMSTORE_TENANT_DIR = ... }, so Kimi CLI, OpenCode, Claude, Cursor etc. all share one memory root even when the program sits on C:\ and projects live on E:. Tenants follow project names, not tools: autosave add E:\proj\B → tenant b (tenant_b.db), regardless of which client you use in that folder. Full walkthrough: docs/user_install.md.

Memory types & half-lives

Each type has a half-life — the time in which an untouched memory's accessibility halves; accessing a memory refreshes it, decay is never deletion, and constraint (0) never decays. Curated default set: semantic, fact, finding, decision, constraint, self, skill (tool/workflow competence), procedural, code (snippets/APIs), plan, goal, progress, question, hypothesis, episodic, lesson, state, observation, snapshot, document, summary, symbolic. Legacy types (instruction, email, log) remain valid for stored history. Half-lives and colors: memstore/core/schema.py.

Documentation

• docs/model_install.md — how AI models/agents should use MemStore.
• docs/user_install.md — non-technical install and daily usage.
• memstore/system_prompt.md — copy-paste prompt for agents (<8k tokens).
• docs/architecture.md (+ the HTML version under docs/architecture.html) — engine internals.

Architecture

• Storage: LMDB with named sub-databases, temporal + tag indices, graph adjacency.
• Primary index: Tantivy full-text over content, tags, memory type, provenance. Pure-Python BM25 fallback when unavailable.
• Retrieval: non-neural-first V2Retriever with metadata filtering, graph expansion, online-learned weight blends, feedback boosting, and sufficiency-aware abstention.
• Trust: provenance/consistency/anomaly trust scoring; trust-gated commits refuse contradictions of high-trust memory.
• Maintenance: lifecycle scheduler applies accessibility decay, quota enforcement, and consolidation summaries.
• Live INDEX: the MCP server describes all tenants in its connection instructions; admin action=index / source_report expose the full map.
• No neural stack: measured dense-fallback quality matched the lexical index, so the dependency was dropped — the product needs no models, CPU or GPU.

Benchmarks & external audit

benchmarks/ contains four independent evaluation suites, all public and reproducible:

• compare_v1_v2.py — head-to-head with the dense v1 stack on a curated 52-query golden set from real tenants. Latest: v2 R@10 = 1.000 (0 misses) vs v1 0.923; median query latency 23 ms vs 626 ms.
• locomo_benchmark.py — official ACL 2024 LoCoMo run (1,986 QA), plus BEAM, adversarial contradiction (F1 1.000), scale/stress, and ingestion throughput harnesses. Scores and methodology in benchmarks/README.md.

python benchmarks/build_golden_set.py
python benchmarks/compare_v1_v2.py
python benchmarks/locomo_benchmark.py
python benchmarks/uat_model_audit.py

License

MemStore is distributed under the MemStore Proprietary License (MPL-1.0) — see LICENSE. It is source-available for inspection and debugging, but it is not open-source: you may use it internally with a valid license, but you may not redistribute it, resell it, sublicense it, or build a competing paid product from it. Update checks and license verification are local-only.