Coding agents lose useful decisions when a session ends. ZestMem gives a team one durable, searchable memory service without storing their transcripts or handing their operational knowledge to a hosted platform.

Agents use two core MCP tools:

agent-a → remember("Production deploys require signed images")
                         ↓
              PostgreSQL + vector search
                         ↓
agent-b → recall("What constrains production deploys?")
          "Production deploys require signed images"

ZestMem is self-hosted, authenticated, and designed for persistent teams using Codex, Claude Code, OpenCode, or any Streamable HTTP MCP client.

ZestMem is not a transcript recorder or a replacement for AgentBus. AgentBus moves durable messages between agents; ZestMem preserves knowledge they should reuse later.

ZestMem is deliberately production-shaped: PostgreSQL 17 with pgvector, a local embedding service, and OIDC authentication. There is no auth-off production mode.

Requirements:

• Docker Compose or Podman
• An OIDC issuer and audience
• Ports 8080 and 15432 available on localhost

git clone https://github.com/jahwag/zestmem.git
cd zestmem
cp .env.example .env
chmod 600 .env

Edit .env and set at least:

POSTGRES_PASSWORD=replace-with-a-long-random-value
OIDC_ISSUER=https://login.microsoftonline.com/YOUR_TENANT_ID/v2.0
OIDC_AUDIENCE=YOUR_MEMORY_API_APP_ID
OIDC_SUBJECT_CLAIM=oid
OIDC_REQUIRED_ROLE=ZestMem.Agent

Entra v2 access tokens put the resource application's bare client-ID GUID in aud; api://YOUR_MEMORY_API_APP_ID/.default is the scope used when requesting the token.

Then start the stack:

docker compose up -d
curl --fail http://127.0.0.1:8080/readyz

The MCP endpoint is http://127.0.0.1:8080/mcp. The memory browser is at http://127.0.0.1:8080/console/.

Prebuilt binaries, checksums, signatures, provenance, and SBOMs are available from GitHub Releases.

OIDC workload tokens and individually revocable native API keys both work on the MCP endpoint. OIDC identities resolve through the exact configured (issuer, sub) or (issuer, oid) binding to a local Principal; ZestMem's Space grants remain the authorization boundary. Give every stable agent its own identity and credential.

The browser console exchanges its entry credential for an opaque, HttpOnly product session. It supports three portable entry modes:

• native Reader API-key exchange;
• standard Authorization Code + PKCE through CONSOLE_OIDC_*; or
• a trusted-edge JWT assertion through CONSOLE_ASSERTION_*.

The trusted-edge assertion authenticates only the human browser surface. MCP clients authenticate directly to the MCP resource. See .env.example and client setup for the complete settings.

Once an operator has created a principal and space, an authenticated MCP client can store a focused memory:

{
  "name": "remember",
  "arguments": {
    "target_space_id": "team-space-id",
    "title": "Production image policy",
    "body": "Production deploys require signed container images.",
    "use_when": "Planning or reviewing a production deployment",
    "kind": "decision",
    "topics": ["deployment", "security"],
    "idempotency_key": "deploy-image-policy-v1"
  }
}

Another agent can retrieve it without knowing the title or exact wording:

{
  "name": "recall",
  "arguments": {
    "query": "What constrains production deploys?",
    "space_ids": ["team-space-id"],
    "limit": 5
  }
}

Recall returns ranked candidates and compact metadata. Full bodies are fetched only when the client explicitly reads a selected memory.

 Codex / Claude Code / OpenCode
               │
      Streamable HTTP MCP
               │
        ┌──────▼──────┐
 OIDC ─►│   memoryd   │◄─ memoryctl (operator)
        └───┬─────┬───┘
            │     │
   PostgreSQL     local embeddings
   + pgvector     (TEI)

• PostgreSQL is the source of truth.
• Embeddings provide candidate retrieval; lexical evidence and metadata remain visible to clients.
• Memories are revisioned, attributable, and scoped to spaces.
• Workspace claims can constrain which agents may access a space.
• Mutation responses do not echo stored memory bodies.

Run the public endpoints behind TLS and keep PostgreSQL, embedding, health, and metrics ports private. The included Caddy configuration serves the MCP endpoint over HTTPS when MEMORY_DOMAIN is configured.

For a remote deployment, tunnel the console rather than exposing it:

ssh -N -L 18080:127.0.0.1:8080 user@memory-host

Then open http://127.0.0.1:18080/console/.

Operational references:

• Security policy
• Support and Orchard Discord
• GitHub backup and restore
• Release verification
• Release process
• Contribution guide

Health endpoints (/healthz, /readyz, and /metrics) are operational surfaces, not public product APIs.

ZestMem is actively developed and already publishes signed multi-platform binaries and container images. Interfaces may continue to sharpen as more agent teams exercise the model.

Licensed under the MIT License.