mirror/nanobot
1
0
Fork 0
mirror of https://github.com/HKUDS/nanobot.git synced 2026-09-02 17:22:06 +03:00
Code Issues Packages Projects Releases Wiki Activity

refactor(agent): make memory summaries cumulative (#5610)

Browse Source 
* refactor(agent): make memory summaries cumulative

Treat the latest session summary as a replacement checkpoint, preserve it through bounded raw fallbacks, and reserve history.jsonl for Dream ingestion.

* fix(agent): preserve cumulative checkpoint context

* fix(agent): preserve memory archive prompt cache

* refactor(agent): state archive prompt positively

* refactor(agent): remove checkpoint version migration

* refactor(agent): summarize full archive context

* test(agent): align cumulative archive prompt assertion

* refactor(agent): clarify memory checkpoint contract
This commit is contained in:
chengyongru
2026-08-31 13:37:20 +08:00
committed by GitHub
parent 6cd7063682
commit bb34b58f47
12 changed files with 476 additions and 514 deletions
Download Patch File Download Diff File Expand all files Collapse all files
docs/configuration.md
+4 -8
View File
@@ -2268,16 +2268,12 @@ When a user is idle for longer than a configured threshold, nanobot **proactivel
How it works:
1. **Idle detection**: On each idle tick (~1 s), checks whether an idle-session scan is due. By default, the full scan runs at most once per minute.
2. **Background compaction**: Idle sessions summarize the older live prefix via LLM and keep the most recent legal suffix (currently 8 messages).
3. **Summary injection**: When the user returns, the summary is injected as runtime context (one-shot, not persisted) alongside the retained recent suffix.
4. **Restart-safe resume**: The summary is also mirrored into session metadata so it can still be recovered after a process restart.
2. **Background compaction**: Older context is summarized while the most recent messages remain available.
3. **Session preservation**: The complete session history remains stored for later inspection and reuse.
4. **Restart-safe resume**: The compacted context remains available after a process restart.
> [!NOTE]
> Mental model: "summarize older context, keep the freshest live turns, **and overwrite the session file with the compact form.**" It is not a full `session.clear()`, but it is a write — not a soft cursor move.
>
> Concretely, auto compact rewrites `sessions/<key>.jsonl` in place: older messages (including their structured `tool_calls` / `tool_call_id` / `reasoning_content`) are replaced by just the retained recent suffix (currently 8 messages), while the archived prefix is preserved only as a plain-text summary appended to `memory/history.jsonl` (or a `[RAW] ...` flattened dump if LLM summarization fails). The original structured JSON of those turns is no longer recoverable from the session file.
>
> This differs from the **token-driven soft consolidation** that fires when a prompt exceeds the context budget: that path only advances an internal `last_consolidated` cursor and leaves the session file untouched, so the raw tool-call trail stays on disk and can still be replayed or audited. If you rely on that trail for debugging or auditing, set `idleCompactAfterMinutes` to `0` and let only the token-driven path run.
> Auto compact shortens the context sent to the model without deleting the session's structured message history.
## Timezone