Context-provider adapters

AtMem normally owns governed retrieval. Context-provider adapters are an optional deployment mode for teams that want Mem0, a LangGraph, or a Pydantic AI agent to remain the context-decision authority while AtMem verifies and delivers the exact result once and records the agent flight.

Installing or starting an adapter does not enable it. The safe sequence is:

1. install exactly one optional extra;
2. initialize and start its private loopback service;
3. register its public key, request-credential references and exact user, agent, and workspace scopes;
4. inspect status and doctor output;
5. explicitly enable that registration.

Disable the registration to return immediately to native AtMem authority.

Mem0

Use Mem0 OSS locally:

python -m pip install 'atmem[mem0]'
atmem provider init mem0-local --kind mem0 --mode oss --port 8788
atmem provider start mem0-local
atmem provider doctor mem0-local

The Mem0 extra installs the verified Python 2.x SDK range mem0ai>=2.0.20,<3. AtMem calls both OSS and Platform search with mandatory filters={"user_id", "agent_id", "app_id"} plus bounded top_k; it never retries without those three scope boundaries. Accepted results retain Mem0's ranked order when the provider runtime builds the signed context proposal.

For Mem0 Platform, export the credential only in the provider process environment and select platform mode:

export MEM0_API_KEY='...'
atmem provider init mem0-cloud --kind mem0 --mode platform --egress hosted --port 8788

Every search carries all three authenticated boundaries in one call: user_id, agent_id, and app_id (the AtMem workspace). The adapter never retries without them. A contradictory returned scope is discarded.

If your application already constructs a configured Mem0 client, expose a zero-argument Python factory and pass --factory package.module:build_client.

LangGraph

python -m pip install 'atmem[langgraph-provider]'
atmem provider init graph-context --kind langgraph \
  --factory myapp.memory_graph:build_graph --port 8789
atmem provider start graph-context

The factory returns a compiled graph or compatible callable. It receives a fresh delegated_request dictionary and a fresh config whose thread_id is bound to the authenticated session and turn. Its final output is closed:

{
    "context_decision": {
        "decision": "inject",
        "items": [{"text": "User likes burgers.", "source_ref": "memory:42"}],
        "source_refs": ["memory:42"],
    }
}

Sync and async graphs are supported. An interrupt or malformed result fails closed and produces no signed authorization.

Pydantic AI

python -m pip install 'atmem[pydantic-provider]'
atmem provider init ai-context --kind pydantic-ai \
  --factory myapp.memory_agent:build_agent --egress local --port 8790
atmem provider start ai-context

The factory returns a configured Pydantic AI agent. Use atmem.provider_adapters.pydantic_ai.proposal_output_model() as its output_type, or return a compatible validated output model. AtMem reads only AgentRunResult.output; free-form model text is not accepted as authority. Mark hosted models with --egress hosted so status and receipt attribution do not imply local processing.

Register and enable exact trust

atmem provider init prints a complete registration command with three placeholders. Replace them using authenticated IDs from your host integration:

atmem delegated register \
  --provider-id mem0-context-provider --provider-version 1.0 \
  --instance-id mem0-local --key-id primary \
  --public-key-file ~/.atmem/providers/mem0-local/public.key \
  --endpoint http://127.0.0.1:8788/v1/delegated-context \
  --request-key-id REQUEST_KEY_ID --request-secret-file /absolute/private/request.key \
  --workspace WORKSPACE_ID --agent AGENT_ID --user USER_ID
atmem delegated status
atmem delegated enable mem0-context-provider:mem0-local
atmem delegated doctor

Provider failure withholds context by default. Native fallback exists only if you deliberately add --native-fallback at registration. AtMem still verifies binding, signature, expiry, replay state, receipt digest, and exact context bytes, and the host still has to confirm what reached model input.

Stop and roll back

atmem delegated disable mem0-context-provider:mem0-local
atmem provider stop mem0-local
atmem provider remove mem0-local --yes

Removal deletes the provider's local key and configuration. It does not delete the external provider's database and does not erase AtMem's historical flight evidence.

Authenticated HTTP and beta migration

Beta 10 requires the shared HMAC request profile. Newly initialized instances generate their own request secret independently of their Ed25519 keys. The init command prints the exact request key ID and secret-file reference in its registration command; use those values instead of the placeholders above.

Unsigned context and health requests are rejected before provider access. Doctor checks authenticated health. An identical HTTP replay is rejected even after restart; signed-response acceptance idempotency is unchanged.

For an old managed instance, stop it, run atmem provider auth-init INSTANCE, and start it again. Attach the returned credential references with atmem delegated set-request-auth PROVIDER:INSTANCE --request-key-id KEY_ID --request-secret-file /absolute/private/request.key, check doctor, then explicitly enable. Legacy enabled registrations without credentials stay blocked until migrated or explicitly disabled. Provider auth-rotate and auth-revoke support bounded overlap without restarting; see the profile for exact commands.

Custom providers such as Storizon must adopt the same HTTP profile and generate their own per-installation request secret. AtMem's managed-provider commands do not configure an external Storizon service. Public demo keys remain test-only.

2.2.6 delegated host compatibility

The maintainer-run acceptance scope is OpenClaw 2026.9.2 and 2026.9.3 with Claude CLI 2.1.236, real local Mem0, Qdrant and Ollama, and dummy data. A separate attributed Storizon rerun used the same OpenClaw versions with Claude CLI 2.1.251, Haiku 4.5, Node 26.5.0 and Linux against the published AtMem and bridge 2.2.6 artifacts.

The CLI sends terminal results on the host tool-event stream when typed completion hooks are absent. The bridge matches those observations to requests and records result digests; requests or model text never create completion evidence. Missing, conflicting or mismatched observations still fail closed. This matrix does not promise every historical/future OpenClaw version or every channel/backend.

See identity configuration, maintainer-run Mem0 evidence, and the attributed Storizon 2.2.6 report.