cc70c3e524
Two-tier memory model (working memory + long-term store) with a unified user namespace across all channels. Addresses four gaps: cross-session forgetting, compaction context loss, no proactive recall, and channel fragmentation. Key design decisions: - user/working namespace written on every compaction (TTL-based expiry) - user/profile + user/patterns as shared identity across channels - Session-start injection before first turn (one-time, idempotent) - Opt-in via memory.user_namespace config; default is unchanged behavior Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Flynn Documentation (Start Here)
This documentation is written to be useful to both humans and AI agents. If you want a fast mental model, read the architecture diagram first.
Reading Order
- Architecture overview (agent-oriented)
docs/architecture/AGENT_DIAGRAM.mddocs/architecture/AGENT_REPO_MAP.mddocs/architecture/GATEWAY_SESSIONS_AND_QUEUE.mddocs/architecture/TYPESCRIPT_MAP.mddocs/architecture/SYMBOL_INDEX.mddocs/architecture/CONTRIBUTOR_MAP.md
- Security boundary (safe-by-default personal agent)
docs/security/SAFE_PERSONAL_AGENT.md
- Gateway API (WebSocket JSON-RPC)
docs/api/PROTOCOL.md
- Tool contracts and policy/execution details
docs/api/TOOLS.md
- Production and performance
docs/deployment/PRODUCTION.mddocs/performance/TUNING.md
- Operations runbooks
docs/operations/OPERATOR_PACK.mddocs/operations/GOOGLE_AUTH.md
Quick Map (One Diagram)
flowchart TD
U[User on chat surface] --> CA[ChannelAdapter]
U --> GW[Gateway WS/HTTP + Web UI]
CA --> RT[Routing + Sessions]
GW --> RT
RT --> OR[AgentOrchestrator]
OR --> NA[NativeAgent tool loop]
NA --> MR[ModelRouter]
MR --> MC[ModelClient -> provider]
NA --> TP[ToolPolicy + ToolRegistry]
TP --> TE[ToolExecutor\nhooks + enforcement + audit]