Skip to content

agent-assets.md

Latest commit

 

History

History
130 lines (104 loc) · 6.33 KB

agent-assets.md

File metadata and controls

130 lines (104 loc) · 6.33 KB
Raw

Governed agent memory assets

LongMemory turns agent work into four reusable project assets without creating a second memory engine:

Input Asset Initial state Typical use
Imported conversations chat_memory candidate Session continuity and prior decisions
Curated procedures skill approved Direct instructions and validation rules
Document connector sync llm_wiki candidate Structured pages, citations, and project knowledge
Repository connector sync code_graph candidate Symbols, callers, callees, and impact analysis

Candidates require an explicit governance update before they can enter an agent loadout. Manually authored Skills are approved because the author supplied their trigger, instructions, validation, and resources explicitly.

Asset model

Every asset has a stable ID and immutable versions. A version records:

  • Type, name, description, owner, source, and readable content_ref
  • Lifecycle: draft, candidate, approved, deprecated, archived, or failed
  • Visibility: private, project, team, restricted, agent, or task
  • Deny-first ACL entries for users, teams, roles, agents, tasks, and frameworks
  • Bindings to agents, tasks, or frameworks
  • Injection mode: direct, summary, tool, or reference
  • Priority, confidence, expiry, labels, payload, and provenance metadata

Lifecycle transitions are validated. Content and policy changes create a new version rather than rewriting history. Visibility grants only read and use; manage, assign, and share require ownership or an explicit ACL.

Loadout assembly

const loadout = await projects.resolveAssetLoadout("longmemory", {
  query: "review the release architecture",
  user_id: "alice",
  team_ids: ["core"],
  roles: ["developer"],
  agent_id: "reviewer",
  task_id: "release-42",
  framework: "codex",
  token_budget: 2_048,
});

Resolution is deterministic and explainable:

  1. Filter by requested type, approved state, expiry, and deny-first access.
  2. Match enabled agent/task/framework bindings.
  3. Reject assets bound to another target.
  4. Rank required bindings, priority, query relevance, and confidence.
  5. Pack structured context under the token budget.
  6. Return every exclusion with a reason.

Selected items include MCP-compatible assistant audience, priority, and last modified annotations. Tool/reference modes keep large Wiki and CodeGraph assets out of the prompt until an agent actually needs them.

Portable agent manifest

const manifest = await projects.buildAgentManifest("longmemory", {
  agent_id: "reviewer",
  framework: "codex",
  query: "review the release",
  user_id: "alice",
  interface_url: "https://agents.example.com/a2a",
});

The versioned manifest includes the authorized loadout, MCP discovery URIs, and the LongMemory extension URI https://longmemory.dev/extensions/memory-assets/v1. When an interface URL is provided it also emits an A2A 1.0-compatible Agent Card containing only the approved Skill capabilities visible to that identity. This is a discovery artifact; LongMemory does not claim to implement the full A2A task protocol.

CLI

longmemory asset list
longmemory asset register --type llm_wiki --name "Architecture wiki" `
  --description "Project architecture" --owner alice --source-type docs `
  --content-ref longmemory://project/current/wiki --status candidate
longmemory asset govern <asset-id> --status approved `
  --agents reviewer --mode tool --priority 0.8
longmemory asset loadout "review architecture" --agent reviewer --framework codex
longmemory agent manifest reviewer --framework codex --query "review architecture"

For complete ACLs, bindings, payload, and metadata, use --input-json or --patch-json.

MCP

  • longmemory_asset_catalog: list/get governed assets or resolve a loadout
  • longmemory_manage_asset: register or govern an asset; blocked in read-only mode
  • longmemory://project/{project_id}/assets: identity-filtered catalog
  • longmemory://project/{project_id}/asset/{asset_id}: one authorized asset
  • longmemory://project/{project_id}/agent/{agent_id}/manifest: portable manifest

MCP identity is fixed by runtime configuration. Tool arguments cannot impersonate another agent or framework. Tool calls are schema-validated and audited.

Design references

The implementation incorporates current ecosystem guidance:

  • MCP resources: application-driven discovery, URI templates, subscriptions, audience/priority/last-modified annotations, and per-resource access checks.
  • MCP tools: strict input validation, access control, sanitization, timeouts, auditing, and human control for sensitive actions.
  • MCP authorization: resource-bound credentials, audience validation, and no token passthrough.
  • A2A 1.0: Agent Cards, capability/Skill discovery, authenticated extended cards, versioned extensions, task/context identity, and authorization scoping.
  • OpenAI Agents sessions: stable session IDs, selective history inclusion, bounded retrieval, shared sessions, and explicit compaction concurrency boundaries.
  • TencentDB Agent Memory: the four-asset product model, human governance, fixed agent bindings, and direct/summary/tool/reference injection modes.