Files
IdeA/.ideai/memory/agent-context-memory-and-profile-handoff.md
Blomios eca2ba95c4 feat(agent): conversation par paire + entrée médiée + pivot terminal/MCP
Coeur inter-agents consolidé et surface front réalignée sur la décision
"terminal natif PTY, pas d'UI chat" (Option 1).

Domaine
- nouveaux modules conversation, mailbox, input, fileguard (ports + types)
- orchestrator/profile/events étendus (conversation par paire, FIFO)

Application / Infrastructure
- orchestrator/service + context_guard : sérialisation FIFO par agent,
  garde RW mémoire/contexte, dispatch ask/reply
- adapters in-memory conversation / mailbox / input / fileguard
- registry session + lifecycle agent durcis (1 agent = 1 session vivante)
- outils MCP idea_* alignés sur le nouveau dispatch

Frontend
- MediatedInput + useAgentBusy : entrée utilisateur médiée par IdeA,
  terminal = vue sortie inchangée
- suppression de la vue chat structurée (AgentChatView) — abandonnée
- adapter input + ports mis à jour

Divers
- .ideai/ : mémoire projet + briefs de cadrage versionnés ;
  requests/ runtime ignoré ; agents projet réels (DevBackend/DevFrontend/QA)

Tests : Rust (domain/application/infrastructure/app-tauri) + front (346) verts.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-12 07:33:04 +02:00

107 lines
5.6 KiB
Markdown

---
name: agent-context-memory-and-profile-handoff
description: Decisions sur l'injection de contexte, la memoire durable, l'etat live et le handoff de profil entre agents IA.
metadata:
type: project
---
# Agent Context, Memory, and Profile Handoff
## Resume
This note captures the current product direction for IdeA around agent context injection, project memory, live state, and profile handoff between AI providers.
See also:
- `idea-product-directives-main-handoff` for product priorities and UX constraints.
- `remaining-work-idea-agent-control-ide` for the current implementation status and remaining work.
## Context Injection
- Agent context must be injected by IdeA at launch time; the model should not be expected to discover `AGENTS.md` or `CLAUDE.md` by itself.
- The existing `contextInjection` architecture is the right mechanism, especially `conventionFile` for providers that support a conventional file in the run directory.
- The current implementation is strongest for `conventionFile`; `flag`, `stdin`, and `env` do not yet receive the same fully-composed IdeA context.
- The `flag` strategy appears fragile with the isolated run directory model because the relative path passed to the CLI may not resolve from the run directory.
## Shared Project Context
- `.ideai/CONTEXT.md` is intended as shared project context.
- It is not created automatically; it only exists if something writes it.
- It should carry active project constraints and contribution rules.
- Example content for `CONTEXT.md`: architectural constraints, workflow rules, and operating conventions that every agent must apply immediately.
## Durable Memory
- `.ideai/memory/` is intended as durable, project-scoped memory shared by all agents of the same project.
- It is not created automatically; it only exists once at least one memory note is saved.
- The durable memory is a knowledge base, not a live activity log.
- It should contain stabilised knowledge such as:
- architecture decisions
- feature summaries to implement later
- user preferences that persist across sessions
- important project facts and references
- Durable memory should stay curated and low-noise.
## Live State Versus Durable Memory
- Durable memory should not be used as a shared real-time state feed for all agents.
- IdeA should distinguish between:
- global context (`.ideai/CONTEXT.md`)
- durable memory (`.ideai/memory/`)
- live operational state (separate store)
- handoff summaries between sessions or profiles
- Real-time work tracking, who-is-doing-what, and transient status should live in a dedicated state mechanism, not in durable memory.
## Memory Consumption by Agents
- The current launcher reads shared project context from `.ideai/CONTEXT.md` if present.
- It also recalls project memory via `MemoryRecall` and injects a `# Memoire projet` section into the convention file.
- This injection currently happens only for `conventionFile` profiles.
- The recalled memory is shared at the project level, but each agent may receive a different subset depending on its persona and recall query.
- Memory recall is computed at launch time and written into the generated context file.
- There is currently no automatic live refresh when `.ideai/memory/` changes during an active session.
## Recommendation for Live Memory Refresh
- Automatic memory refresh could be useful, but it should be explicit and controlled.
- If IdeA wants agents to benefit from memory changes while they are active, it should regenerate their effective context when needed instead of treating durable memory as a constantly streaming log.
- For PTY agents, no automatic reread exists today.
- For structured Claude/Codex sessions, each turn relaunches the CLI, but the generated convention file is not automatically rewritten when durable memory changes.
## Profile Handoff and Session Continuity
- A direct native session transfer from Claude to Codex is not the right mental model.
- The correct model is continuity of work state, not native provider-session continuity.
- IdeA should persist:
- a canonical conversation log
- a cumulative handoff summary
- the agent state
- the provider conversation id when useful
- On provider swap, IdeA should launch the new profile with:
- regenerated project context
- regenerated durable memory recall
- the current agent persona
- a handoff summary plus recent transcript
- The handoff summary should not be created only at the moment of swap.
- IdeA should maintain summaries incrementally or at checkpoints so that a swap is still possible when the current provider is near a token or session limit.
## Agent Ability To Write Durable Memory
- Agents should be allowed to promote important knowledge into durable memory.
- This should not rely on the agent guessing the capability.
- IdeA should expose the capability explicitly through tools or commands and should inject a clear rule explaining when an agent may save durable knowledge.
- This write ability should be constrained to stable, high-value information, not transient state.
## Practical Classification Rule
- Put immediate project rules and operating constraints in `.ideai/CONTEXT.md`.
- Put stable, reusable project knowledge in `.ideai/memory/`.
- Put current activity and coordination state in a separate live-state mechanism.
- Put cross-session or cross-profile recovery material in a handoff/session layer.
## Current Product Direction
- Keep `CONTEXT.md` for project rules.
- Keep `.ideai/memory/` for curated durable knowledge.
- Introduce a separate live-state mechanism if agents must stay aligned on in-progress work.
- Introduce persistent conversation logs and incremental handoff summaries to support profile swaps such as Claude to Codex.