MCP memory server for Claude Code and Codex

Memory for AI coding agents that you can hold accountable.
Keep decisions, fixes and project context between sessions, and inspect what was retrieved.
Runs locally by default. No account, no API key, no server to manage.

Sovereign is what it is today. Everything runs on your machine: a local SQLite file by default, or PostgreSQL + pgvector if you prefer. No LLM in the retrieval loop, and nothing leaves localhost unless you configure an integration that does. Your project's memory is a file you own and can delete.

Cross-platform is how it is built. One stdio MCP server and the same 52 tools on Claude Code, in the Claude Desktop bundle, under Claude Cowork, and on every local stdio MCP host listed in the table below. What differs per host is stated there, not discovered after install.

Eco-responsible is what we are aiming at. Work that never reaches a datacenter is work nobody has to power, and an agent that finds the right context first time re-reads fewer files. We hold that intent to the Green Software Foundation's SCI method, and we publish no CO₂ or energy figure, because we have not measured one. What we do and do not claim ↓

36 neuroscience mechanisms · 52 memory tools · 9 lifecycle hooks · a self-curating per-project wiki — all local, all open-source, MIT.

Install

Claude Code — add the marketplace and install the plugin:

claude plugin marketplace add cdeust/Cortex
claude plugin install hypermnesia-mcp

Claude Desktop — download hypermnesia-mcp.mcpb from Releases and open it, or use Settings → Extensions. The bundle carries the tools but no hooks; the MCPB format has none.

Claude Cowork is detected automatically (CLAUDE_ENVIRONMENT=cowork) and uses the local SQLite store. No PostgreSQL required.

Any other stdio MCP host (Codex, Gemini CLI, Cursor, Windsurf, VS Code) launches the same server and gets the same tools. The per-host matrix and launch commands are in Every other MCP host below. Codex has a native package: docs/codex-plugin.md. WSL, TLS client certificates and corporate proxies are covered in docs/deployment-scenarios.md.

The first use creates a local SQLite store under ~/.claude/methodology/. Models are downloaded once when needed and then run offline. The embedding and reranking model files are both fetched on first use. Optional integrations, remote PostgreSQL, and OTLP telemetry use the network only when explicitly configured. PRIVACY.md lists the exact scope.

An existing PostgreSQL install is never silently downgraded: the installer detects a configured DATABASE_URL, a prior backend marker, or a reachable local cortex database and keeps it across updates.

claude plugin uninstall cortex
claude plugin install hypermnesia-mcp

claude plugin uninstall cortex-viz@cortex-plugins
claude plugin marketplace update cortex-plugins
claude plugin install hypermnesia-mcp-viz@cortex-plugins

Keep context useful

Across sessions, agents need to remember decisions, bring prior fixes back when a similar problem returns, and show you which sources support a memory so you can correct it. Cortex keeps that context available while making its status visible.

Cortex does this with local quality checks: what is written, whether its references resolve, what happens when a decision changes, and what can fade over time.

What gets in

A write passes a local novelty check (the implementation calls it a predictive-coding gate) against what is already stored. Novel content is written; a near-duplicate is merged into the memory it restates rather than filed beside it.

// Illustrative project decision:
remember({ content: "Keep session state in Redis so TTL expiry is handled consistently." })
// → { stored: true, action: "stored" }

Deliberate writes are never rejected for being unsurprising. Unattended capture is, which is what keeps automatic capture from burying the memories you meant to keep.

Whether it can be checked

Every memory is graded at write time, locally, with no network call. The grade is not a confidence score: it is whether the claims carry references that resolve on this machine.

// → provenance: { grade: "unverifiable",
//                 reason: "dead_refs: deps/numpy/_core/_multiarray_umath.cpython-313-darwin.so",
//                 hint: "1 of 9 checkable reference(s) could not be resolved" }

That memory named a file that no longer existed, so it was stored and labelled unverifiable instead of being silently presented as verified. Rewritten against paths that resolve, the same memory grades verified. A recalled memory tells you which kind it is; a verified grade still means that the references resolve locally, not that the claim has been independently proven true.

When it turns out wrong

Corrections supersede rather than overwrite. The new memory records what it replaces, the old one is demoted in recall, and the chain stays readable.

remember({ content: "...", supersedes_id: 4360411 })
// → { action: "superseded", memory_id: 4360412, superseded_id: 4360411 }

What fades

Memories carry heat that decays unless replay reinforces them, and episodic traces can consolidate into semantic ones. A specific debugging session may compress to the principle it taught; the commands can fade while the lesson survives. This lifecycle is designed to keep the store useful as it grows, though it is not a promise of a fixed size or guaranteed semantic compression.

What it feels like in use

Here is an illustrative workflow: decisions, prior fixes, and source checks becoming useful again.

Monday. An hour debugging a webhook handler ends in a race condition: TTL expiry firing between the auth check and the permission lookup. You agree on a fix, implement it, close the session.

Thursday. In another session, a user reports intermittent logouts. Cortex surfaces relevant prior analysis, the Redis decision, and the TTL lesson when their content matches the new work.

Three weeks later. The sessions can consolidate into a pattern about authentication and TTL-based caches; some details may fade while the principle remains useful.

In Claude Code that is automatic: nine lifecycle hooks inject context at session start, recall per prompt, capture as you work, checkpoint before compaction, and run a per-project wiki that curates itself. In any other stdio MCP host you call the same 52 tools yourself, or 55 when the optional ai-architect-mcp-codebase and ai-architect-mcp-spec integrations are present.

Does the retrieval work

Measured against a published benchmark, retrieval only. No LLM reader in the loop: the question is whether the right memory surfaces, not whether a model can write a good answer from it.

LongMemEval: 500 human-curated questions buried in about 40 sessions of history.

Both are single runs: n=500, clean database, consolidation disabled, retrieval only.

v4.14.1, 2026-07-14: artifact JSON; code SHA. This is the run the ablation campaign in Verification was built around.

v4.20.0, 2026-09-09: artifact JSON and its manifest; code SHA, dirty=false. A single run of the LongMemEval-S leg alone (benchmarks/reproduce.sh --only longmemeval --no-ablation) in an isolated ephemeral PostgreSQL container, reranker loaded, consolidation disabled. Against v4.14.1 the change is 0.4 points of Recall@10 and 0.012 of MRR. The run's own floor check reports Recall@10 within the 0.005 tolerance of the July floor (0.982) and MRR 0.0093 below its floor (0.914), which the script treats as non-blocking by design; docs/agent-guidance.md records that main no longer clears those floors and that the release gate is --no-regression against origin/main. The same tree has no LoCoMo or BEAM figure yet.

Reproduce with benchmarks/reproduce.sh, which runs in an isolated ephemeral container, never against a live store.

Recall@10 is the share of questions whose answer-bearing session appears in the first ten retrieved sessions. MRR (mean reciprocal rank) rewards finding that session near the top. These numbers describe retrieval only; they do not measure whether an LLM writes a correct answer.

Retrieval fuses five signals through weighted reciprocal-rank fusion, then reranks with a cross-encoder: vector similarity, full-text search, trigram match, heat and recency. LoCoMo and BEAM results, the ablations and the floor gates are in benchmarks/.

Storage

SQLite by default. PostgreSQL is one configuration field, worth it for very large stores or a database shared across a team.

bash <plugin-dir>/scripts/install-plugin.sh --postgres

Three hook enrichments are PostgreSQL-only and degrade to silent no-ops on SQLite. Session banners, auto-recall, auto-capture, checkpoints and every memory tool work on both.

Every other MCP host

The server is host-agnostic. Any host that can launch a stdio process gets the full tool surface on the default SQLite store. What is not portable are the nine lifecycle hooks, which are Claude Code plugin machinery; the server never imports or requires them at startup.

On Claude Code memory is ambient: hooks capture and inject automatically. On every other host memory is tool-driven: the agent stores and retrieves when instructed, and nothing happens between prompts.

The launch command on every host is the PyPI package. The [sqlite] extra enables sqlite-vec vector search; without it the store still works, with vector search disabled.

uvx --from "hypermnesia-mcp[sqlite]" hypermnesia-mcp

Gemini CLI ships as an extension (gemini-extension.json is in this repository):

gemini extensions install https://github.com/cdeust/Cortex

Codex and ChatGPT desktop have a native plugin with a 10-tool lean surface. Pre-install the package once so the plugin's first uvx handshake reuses the local uv cache instead of spending its startup budget downloading a Python environment:

uv tool install "hypermnesia-mcp[sqlite]"
codex plugin marketplace add cdeust/Cortex
codex plugin add hypermnesia-mcp-codex@cortex-codex-plugins

The direct fallback registers the executable without the plugin:

codex mcp add cortex --env CORTEX_MEMORY_STORE_BACKEND=sqlite -- hypermnesia-mcp

The host boundary, the measured startup ceiling and the public-directory requirements Cortex deliberately does not claim are in docs/codex-plugin.md.

Green software engineering

Cortex runs a standing efficiency programme, gated by the same evidence rule as the retrieval work: no unsourced efficiency claim ships. Waste is treated as a defect with a reproduction, not as a virtue to advertise.

The measurement harness — and what it does not establish

benchmarks/energy/ implements the Green Software Foundation SCI specification: operational emissions O = E × I, embodied allocation M = TE × TS × RS, reported per functional unit. For the embedding path the functional unit is 1000 model input tokens, counted from the tokenizer's own attention_mask — never estimated from characters.

Read benchmarks/energy/README.md before quoting anything from it. Its own first paragraph is the important one: the automated fixtures exercise arithmetic and failure paths, they do not measure device energy and do not establish an energy improvement. Further, by design:

• No default carbon factors. --carbon-intensity (gCO2eq/kWh) and --embodied (gCO2eq/s, an already allocated rate) are mandatory operator inputs, validated before any model import. The harness records the values and their units; it does not vouch for their provenance. You supply the region, observation period, lifecycle assessment and reservation assumptions.
• A stated boundary. raw_system_energy_j is the sensor's combined CPU+GPU+ANE estimate. It is neither wall-plug energy nor a complete device SCI score: memory, storage, screen, power-supply losses, model warm-up and token counting are all excluded.
• Artifacts or it did not happen. A successful run preserves results.json, a MANIFEST.json of commit and source hashes, and the exact analyzed powermetrics.txt snapshot.

No energy results are committed to this repository. That is deliberate: a figure measured on one operator's machine, region and duty cycle is not a property of the software, and publishing it as one would be the drift this programme exists to prevent.

What has actually shipped

Efficiency work lands as ordinary reviewed PRs. Two workstreams are merged:

The hook work is the load-bearing one, because hooks run on every tool event. Deferring the handler/store stack keeps hook boot at ~0.05 s against ~0.6 s for the full registry import (measured 2026-07-28; the constant is cited in mcp_server/hooks/auto_recall.py at its call sites, per the no-invented-constants rule).

Demand reduction is the primary lever

The largest efficiency term in an LLM-assisted workflow is not this server's own CPU — it is the tokens a model must process because the right context was not found the first time. That makes retrieval quality an energy property, and it is why the benchmark tables above and this section are the same programme: response_budget.py bounds a payload and keeps ids so truncation stays resumable, the reranker degrades to first-stage scores rather than fetching a model, and CORTEX_RERANKER_OFFLINE=1 refuses the download outright.

This paragraph is a design rationale, not a measurement. Cortex publishes no token-savings or CO2 figure for end-to-end agent sessions, because it has not measured one.

Verification

The v4.14.1 figures above are backed by a per-mechanism ablation campaign — full n, single-seed, with code SHAs, dirty flags, manifests, and per-row JSON preserved; the v4.20.0 figures are a single measurement without one:

• LongMemEval-S, 17 rows, n=500 — docs/benchmarks/e1-v3-results.md. Per-mechanism deltas at the calibrated equilibrium + category-specialization analysis.
• LoCoMo, 14 rows, n=1986 — docs/benchmarks/e1-v3-locomo-results.md (pre-fix) and docs/benchmarks/e1-v3-locomo-results-post-fix.md (post plasticity result-shape fix). Two-baseline design (NO_CONSOLIDATION / WITH_CONSOLIDATION).

The full per-mechanism evidence lives in the thermodynamic paper (§6.3); the BEAM decay dose-response (§6.4) documents a re-scoped negative result after a dirty-store confound was caught and traced. Thermodynamic Memory vs. Flat-Importance Stores (PDF, 34 pages) · Stage-Aware Context Assembly (PDF, 39 pages).

Under the hood

The mechanisms above are implemented as 36 system mechanisms spanning encoding, consolidation, retrieval and forgetting. Each is cited to published work and exposed as a live system vital. The bibliography is the check: its entry count is what the references badge reports, and a gate fails the build if the two disagree.

Clean Architecture, concentric layers: server → handlers → core ← shared, and infrastructure → shared. Core is pure and testable without mocks. docs/agent-guidance.md is the map; docs/mcp-tools.md is the tool reference.

Limits worth knowing before you install

• The automatic behaviour is Claude Code plugin machinery. Elsewhere you call the tools yourself, and the host table above says exactly what is missing where.
• SQLite fusion is in-process and unindexed. Fine at personal scale, slower at very large one.
• The retrieval scores above are retrieval-only. They say nothing about answer quality.
• Provenance grading is local and structural. It checks that a reference resolves, not that a claim is true; a DOI or arXiv link is never auto-verified.
• No energy or carbon figure is published, for the reasons stated above.
• First use downloads both the embedding and reranking model files. Optional integrations, remote PostgreSQL and OTLP telemetry add network activity only when explicitly configured; see PRIVACY.md.

Security

Runs 100% locally — MCP over stdio, the storage backend (SQLite file or PostgreSQL on localhost) never leaves your machine (the optional hypermnesia-mcp-viz companion binds its server to 127.0.0.1). No data leaves your machine. SafeSkill scan: 94/100 (code 97, content 88 — docs/safeskill-report.json).

Privacy Policy

Cortex is local-first: your memories, conversations, and profiles stay on your machine — stored in a local SQLite database (~/.claude/methodology/memory.db) by default, or in a PostgreSQL database you control. Cortex sends no memories, content, or telemetry to the author, Anthropic, or any third party. The only outbound network activity is a one-time download of open-source embedding/reranking models from Hugging Face (model files only), plus any integrations you explicitly configure. Full policy: PRIVACY.md.

Support

• Issues & bug reports: GitHub Issues
• Security disclosures: see SECURITY.md
• Contact: admin@ai-architect.tools

Development

pytest                                # full suite; assets/badge-tests.svg carries the current count
ruff check . && ruff format --check . # lint and format, both enforced in CI
python scripts/check_doc_claims.py    # advertised counts must match the repo
python scripts/check_craftsmanship.py # file and method caps, layer whitelist, sourced constants

CONTRIBUTING.md describes the gates a change has to clear. GOVERNANCE.md says who decides and what happens if the maintainer stops. docs/ROADMAP.md says where the project is going, and docs/ASSURANCE-CASE.md states the security argument and its limits. CHANGELOG.md is the complete release history.

License

MIT — see LICENSE.

This software is the independent work of Clément Deust. It was developed outside any employment relationship and is not affiliated with, endorsed by, or owned by any past or present employer. It is part of the ai-architect ecosystem (zetetic-team-subagents, ai-architect-mcp-codebase, ai-architect-mcp-spec).

The neuroscience and information-retrieval algorithms encoded in this software are derived from published academic work cited in docs/papers/bibliography.md and inline in the source via # source: annotations (Friston on predictive coding, Anderson & Lebiere on rate-distortion forgetting, Nader et al. on retrieval-induced lability, McClelland et al. on consolidation, and others). The MIT license covers this implementation; it does not assert ownership over the underlying mechanisms, which remain attributable to their original authors and publications.

Citation

The paper PDFs on main are the canonical artefacts (arXiv IDs forthcoming, endorsement in progress):

@software{cortex2026,
  title={Cortex: Persistent Memory for Claude Code},
  author={Deust, Clement},
  year={2026},
  url={https://github.com/cdeust/Cortex}
}

@unpublished{deust2026thermodynamic,
  title={Thermodynamic Memory vs. Flat-Importance Stores:
         Why Long-Term Retrieval Collapses Without Decay},
  author={Deust, Clement},
  year={2026},
  note={arXiv ID forthcoming, endorsement in progress},
  url={https://github.com/cdeust/Cortex/blob/main/docs/arxiv-thermodynamic/main.pdf}
}

@unpublished{deust2026context,
  title={Stage-Aware Context Assembly for Long-Context Memory Retrieval},
  author={Deust, Clement},
  year={2026},
  note={arXiv ID forthcoming, endorsement in progress},
  url={https://github.com/cdeust/Cortex/blob/main/docs/arxiv-context-assembly/main.pdf}
}