Session Log Outside the Context Window

An append-only, queryable session or event log held outside the context window keeps compaction decisions recoverable and lets a crashed harness resume from its last event.

Evidence: strong

If the conversation in the context window is the only record of an agent's session, two things go wrong. Every compaction destroys information that cannot be recovered, and a crash of the harness process loses the session or leaves someone to reconstruct it by hand.

Several providers have converged on keeping a durable log outside the window. Anthropic's Managed Agents design virtualises the session as an append-only event log, separate from both the harness loop and the sandbox. The harness writes events with emitEvent and reads slices with getEvents(). Context transformation stays in the harness, so compaction decisions remain recoverable. A crashed harness is replaced and resumes from the last event using wake(sessionId) and getSession(id). Google's Agent Development Kit treats the session as a durable log of typed events from which the working context is compiled. Microsoft's Agent Framework persists history after each model call, and OpenAI's Agents software development kit (SDK) has sessions as a persistence layer. OpenClaw keeps the full history on disk when it auto-compacts, and Hermes stores sessions in SQLite with full-text search, keeps lineage across compressions and exposes a session_search tool to recall older sessions. Meta's Muse Code writes each session as an append-only JSON Lines (JSONL) log of every model call, tool run and approval, resumes by appending to the same log, and exports it byte-deterministically so audits can be compared; its cookbook also turns curated slices of these logs into CI regression fixtures, described in deterministic session replay in CI.

The log is not memory in the agent memory tiers sense: nothing in it is retrieved into context by default. It resembles a memory stream in being append-only, but its job is fidelity and recovery rather than recall, and it is to an agent session what a knowledge activity journal is to a knowledge base. It is the storage half of context as a compiled view, the "session" leg of brain–hands–session decoupling, and the agent-level counterpart of journal-then-replay durable execution.

Two practical consequences follow. An agent diagnosing another agent should read that agent's durable logs and session store rather than any compacted transcript, because the raw record is what supports diagnosis (raw traces over summaries). And any long-running agent should keep its own append-only log outside its container and its context window, so a crash or restart resumes the task rather than starting it again.

Sources

  • Anthropic Engineering, 8 Apr 2026 — https://www.anthropic.com/engineering/managed-agents
  • Google Developers Blog, 4 Dec 2025 — https://developers.googleblog.com/architecting-efficient-context-aware-multi-agent-framework-for-production/
  • OpenClaw — https://docs.openclaw.ai/concepts/compaction
  • Nous Research — https://hermes-agent.nousresearch.com/docs/developer-guide/architecture
  • Microsoft Learn — https://learn.microsoft.com/en-us/agent-framework/concepts/harness
  • OpenAI, Python SDK docs — https://openai.github.io/openai-agents-python/
  • Nous Research, session_search — https://hermes-agent.nousresearch.com/docs/user-guide/features/memory
  • Meta Model API cookbook, Building with Muse Code, undated (retrieved 3 Oct 2026) — https://dev.meta.ai/docs/cookbook/audit-agent-sessions