mem9

Why Separate Notebooks Per Tool Stop Working

Local prompt files and tool-specific notebooks fragment context. An agent can know something in one runtime and lose it in another. mem9 moves storage, search, ingest, and policy behind a shared service so thin integrations can access the same memory space.

Example memory requests:

Remember that the production API uses the EU region and changes require a rollback note.

Recall decisions about authentication made in previous sessions before editing this service.

Automatic Recall, Smart Ingest, and the Five Memory Tools

The core API flow supports store, search, get, update, and delete. Smart ingest can extract durable facts when an LLM provider is configured; raw ingest remains available as a fallback. Hybrid recall can combine semantic and keyword search, and the dashboard helps inspect stored information.

Automatic recall is not automatic truth. Agents can store stale or incorrect facts, so important memories need source context, ownership, and update rules.

Compare J-Space for evidence-bound task state and Vibe-Skills for Skill orchestration.

The Same Memory, Shared With Claude Code and Codex Too

Set up shared memory deliberately:

1. Choose the hosted API or a self-hosted mem9 server.
2. Create a memory space and keep its API key private.
3. Install the integration for each approved runtime.
4. Write a test memory with clear provenance.
5. Recall it from a second runtime and inspect the result.

Example output contract:

Store this decision with source, timestamp, owner, scope, and replacement condition. Return the memory ID and normalized fields.

Shared memory increases reuse and also expands the effect of a bad write. Apply least privilege and separate spaces when teams or environments should not share context.

Installing the mem9 DeepSeek Harness Bundle

Follow the official mem9 DSH integration guide. Set the credential, install the package into the intended profile, and inspect the composed configuration:

export MEM9_API_KEY="<mem9-api-key>"
dsh plugin --profile <profile> add @mem9/dsh-plugin
dsh --profile <profile> --dump-config

For a self-hosted origin, override the complete mem9 row in that profile's cordis.patch.yml; the endpoint belongs in config.apiUrl, not an export MEM9_API_URL line:

- id: mem9
  name: '@mem9/dsh-plugin'
  config:
    apiKeyEnv: MEM9_API_KEY
    apiUrl: "<self-hosted-mem9-origin>"

Replace <self-hosted-mem9-origin> with the HTTPS base URL of your own self-hosted mem9 server before using this configuration.

Repeat any other non-default keys that must remain active because the config object is replaced as a whole.

Memory-integration status verified 2026-09-28: the DeepSeek Harness project remains in developer preview and expects compatibility-breaking changes. Dump and inspect the profile configuration after an upgrade.

Publish Proven Memory Practices Through EvoMap

mem9 stores and retrieves memory; EvoMap is separate infrastructure for organizing verified agent assets. After provenance, ownership, access, and replacement rules are approved, a team can explicitly package the memory policy. The EvoMap Wiki explains the asset model.

FAQ

Is mem9 only for DeepSeek Harness?

No. It supports several agent runtimes and direct HTTP clients through the same server model. Vibe-Skills can coordinate workflows that use more than one local Skill.

Does smart ingest require an LLM configuration?

Smart ingest falls back to raw ingest when its LLM key is not configured. The stored result still needs provenance and lifecycle rules.

Can mem9 be self-hosted?

Yes. A Go server and multiple database backends support self-hosting. Self-hosting adds deployment and security responsibilities.

Are recalled memories always correct?

No. Memory retrieval returns stored information, which still needs provenance checks and lifecycle management. Remove or replace stale facts when their scope changes.

Related Pages

• Coordinate Skills with Vibe-Skills
• Protect long tasks with J-Space
• Collect browser evidence with BrowserSkill

Reuse a Governed Memory Policy

Turn a tested memory policy into reusable experience while keeping provenance, ownership, access boundaries, and replacement conditions explicit.

Explore EvoMap knowledge assets Compare Vibe-Skills orchestration