Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    strata-memory: Persistent MCP Memory for AI Agents

    1

    Most agents start every session with no memory, and the ones that do remember often paste everything they know back into the prompt.

    Free

    Security scanned
    strata-memory: Persistent MCP Memory for AI Agents

    strata-memory: Persistent MCP Memory for AI Agents

    Example session with this skill installed

    Remember that I prefer dark mode in VS Code with the Monokai Pro theme. Also note that I maintain strata-memory. Next time I ask about my editor setup, use that.

    • Read your context and instructions
    • Compiled the strata-memory

    Result: (real run, 2026-10-28 Asia/Shanghai; sanitized — no keys)

    strata_init (personal): status ok. Embedding siliconflow/BAAI/bge-m3. CBT passive.

    strata_doctor: status

    ok, healthy

    true, embedding_api_key usable=true (provider siliconflow, model BAAI/bge-m3). Recommendation: "All clear."

    commit_memory ×2 (session sess_demo_2026-10-28, Quality Kernel accepted both):

    memory_idtypeconfidencesummary
    39ad30501084488fuser_preference / L00.92User prefers dark mode in VS Code with the Monokai Pro theme.
    a9bd73856b524d5ffactual_truth / L00.95User maintains strata-memory, an open-source MCP memory hub for AI agents.

    recall_context query "What is my VS Code theme preference?" (depth deep): status

    ok, "Recall funnel returned 2 cards (~44 tokens).", fusion vector_hits=2, rerank multi_signal_v3. Top card id 39ad30501084488f, score

    1.0 (recency 1.0, relevance 1.0, importance 0.75, why "L0 core profile").

    expand_memory_detail (user_id vincy_demo, memory_id 39ad30501084488f): status

    ok, "Full detail loaded under progressive disclosure.", fact_claim matches the preference, confidence

    0.92, importance

    0.75, authority user, tags editor/theme/vscode, created 2026-10-28 10:21 Asia/Shanghai, truncated false.

    Context stayed small: short cards first, full text only for the one id that mattered.

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Most agents start every session with no memory, and the ones that do remember often paste everything they know back into the prompt. strata-memory is an open-source MCP memory server built to fix both problems: it keeps durable facts in a local SQLite database and gives the agent a careful, two-step way to recall them.

    This free skill teaches your agent how to install strata-memory, connect it to your host, and use it correctly, so memory stays accurate, scoped, and small in context.

    What it does

    • SQLite as the single source of truth. Memories, audit records, and trajectories live in SQLite. The Chroma vector index can be rebuilt from it at any time, and Markdown is a read-only projection for humans.
    • Quality-checked writes. commit_memory passes every fact through a Quality Kernel and CBT middleware that reject passwords, API keys and tokens, pure venting, fuzzy time words such as "yesterday", and guesses stated as fact.
    • Progressive recall. recall_context uses hybrid search (vector + SQLite FTS5, fused with RRF) and returns short cards with an id, summary, and score. expand_memory_detail fetches full text only for the ids that matter.
    • L0–L3 tiers with typed TTL for facts, preferences, procedures, and episodes, plus strata_digest for score- and TTL-based demotion.
    • Scratch to durable. Session notes stay in a scratch buffer until promote_session confirms them.
    • Isolation. Every record is scoped by tenant_id, user_id, and session_id; expanding across those boundaries fails.
    • Maintenance tools. strata_doctor (index consistency and key health), strata_hygiene (expired, duplicate, and secret-residue scan), strata_stats (L0–L3 token watermarks), strata_project (Markdown export), and strata_rebuild_index (requires confirm=true).

    How it works

    1. The agent checks for Python 3.11+, uv, and git, then clones and installs the server from GitHub.
    2. You set embedding credentials yourself (STRATA_API_KEY / STRATA_EMBEDDING_API_KEY, or skip for local / loopback). The skill tells the agent never to ask for the key in chat.
    3. Optionally set STRATA_EMBEDDING_PROVIDER, STRATA_EMBEDDING_BASE_URL, STRATA_EMBEDDING_MODEL, and STRATA_EMBEDDING_DIMENSION to leave the SiliconFlow default.
    4. The agent registers the server using the included templates for Claude Code, Cursor, Claude Desktop, or Hermes.
    5. It runs strata_init once and strata_doctor to confirm everything is healthy.
    6. In each session it recalls first, expands only what it needs, commits vetted facts, and promotes scratch notes when you confirm them.

    What's in the package

    • SKILL.md: setup checklist, tool map, write and recall examples, and safety rules for the agent
    • references/claude-code.mcp.json, references/cursor-mcp.json, references/claude-desktop.json, references/hermes-config.yaml: ready-to-edit host configs (folder name is references/, as shipped on main)

    Where it stops

    • The skill does not bundle the server; it installs it from the public repository.
    • Embeddings are configurable (default SiliconFlow BAAI/bge-m3). Also supported: openai, openai_compatible (any OpenAI-compatible POST /embeddings, including Ollama/llama.cpp/TEI/vLLM on loopback), and local (in-process sentence-transformers; uv sync --extra local). Non-local providers send memory text and recall queries to the configured endpoint; local keeps text on-device. Memories themselves are stored locally under STRATA_PALACE.
    • It never stores secrets and never rebuilds the index without your confirmation.

    Open source

    strata-memory is MIT-licensed and the core stays open source: https://github.com/SherryLi-vc/strata-memory. Issues and pull requests are welcome.

    • Embedding provider is configurable (siliconflow default, also openai, openai_compatible, local). Changing provider or dimension usually requires strata_rebuild_index(confirm=true). Non-local sends text to the configured endpoint.
    • Python 3.11 or newer is required.
    • Installs from source (git clone + uv); the strata-memory-mcp package is not on PyPI yet.
    • The Hermes wrapper needs STRATA_DIR set to your own clone.
    • Version 2.x renamed the 0.2.x tools (memorize → commit_memory, wake_up/search → recall_context + expand_memory_detail); a migration command is included in the repo.

    Changelog: Skill v1.0.0 first Agensi release; server v2.1.0 multi-provider embeddings.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 20 days ago

    • Free with an Agensi account

    Requires

    Python 3.11+
    uv (Python tool runner)
    git
    Python

    Permissions

    Terminal / Shell
    Read Files
    Write Files
    Network Access
    Environment Variables

    Shell runs git clone, uv sync, uv run strata-memory-mcp, and host MCP registration. Write covers host MCP config and the palace folder under STRATA_PALACE. Network is GitHub (clone), PyPI via uv, and the configured embedding endpoint (default SiliconFlow; or OpenAI / compatible / loopback). Reads STRATA_API_KEY, STRATA_EMBEDDING_API_KEY, STRATA_PALACE, and STRATA_EMBEDDING_*.

    Listed20 days ago

    What's inside

    Frequently Asked Questions

    Yes. The skill is free and the server is MIT-licensed open source.

    Locally under STRATA_PALACE (default ~/.strata/palace). With non-`local` embeddings, only text sent for embedding leaves your machine (to the endpoint you configured). With `local`, embedding text stays on-device.

    No. SiliconFlow BAAI/bge-m3 is the default. Set `STRATA_EMBEDDING_PROVIDER` to `openai`, `openai_compatible`, or `local`, plus `STRATA_EMBEDDING_BASE_URL` / `STRATA_EMBEDDING_MODEL` / `STRATA_EMBEDDING_DIMENSION` as needed.

    Yes. Point each agent at the same server and palace; records are scoped by tenant_id, user_id, and session_id.

    Run strata_doctor. If needed, run strata_rebuild_index with confirm=true; SQLite stays the source of truth.

    No. The Quality Kernel rejects secrets, and strata_hygiene scans for any residue.

    New in LLM Ops