Include persisted turn timestamps when assembling LLM prompts so relative-date references like yesterday and today have concrete anchors.
Made-with: Cursor
#3412 stopped the headline raw_archive bloat but left four adjacent leaks
on the same pollution chain:
- archive() success path appended uncapped LLM summaries to history.jsonl,
so a misbehaving LLM could re-open the #3412 bug from the happy path.
- maybe_consolidate_by_tokens did not advance last_consolidated when
archive() fell back to raw_archive, causing duplicate [RAW] dumps of
the same chunk on every subsequent call.
- Dream's Phase 1/2 prompt injected MEMORY.md / SOUL.md / USER.md and
each history entry without caps, so any legacy oversized record (or an
unbounded user edit) would blow past the context window every dream.
- append_history itself had no default cap, leaving future new callers
one forgotten-cap-away from the same vector.
Changes:
- Cap LLM-produced summaries at 8K chars (_ARCHIVE_SUMMARY_MAX_CHARS)
before writing to history.jsonl.
- Advance session.last_consolidated after archive() regardless of whether
it summarized or raw-archived — both outcomes materialize the chunk;
still break the round loop on fallback so a degraded LLM isn't hammered.
- Truncate MEMORY.md / SOUL.md / USER.md and each history entry in Dream's
Phase 1 prompt preview (Phase 2 still reaches full files via read_file).
- Add _HISTORY_ENTRY_HARD_CAP (64K) as belt-and-suspenders default in
append_history with a once-per-store warning, so any new caller that
forgets its own tighter cap gets caught and observable.
Layer the caps by scope: raw_archive=16K, archive summary=8K,
append_history default=64K. Tight per-caller values cover expected
payloads; the wide default only catches regressions.
Tests: +9 regression tests covering each fix. Full suite: 2372 passed.
Made-with: Cursor
Root cause: when consolidation LLM fails, raw_archive() dumped full message
content (~1MB) into history.jsonl with no size limit. Since build_system_prompt()
injects history.jsonl into every system prompt, all subsequent LLM calls exceeded
the 200K context window with error 1261.
Additionally, _cap_consolidation_boundary's 60-message cap caused consolidation
to get stuck on sessions with long tool chains (200+ iterations), triggering
the raw_archive fallback in the first place.
Three-layer fix:
- Remove _cap_consolidation_boundary: let pick_consolidation_boundary drive
chunk sizing based solely on token budget
- Truncate archive() input: use tiktoken to cap formatted text to the model's
input token budget before sending to consolidation LLM
- Truncate raw_archive() output: cap history.jsonl entries at 16K chars
Move the non-int cursor guard out of the two consumer sites and into a
shared ``_iter_valid_entries`` iterator so the invariant lives in one
place. Closes three gaps left by the original fix:
* ``bool`` is now rejected — ``isinstance(True, int)`` is ``True`` in
Python, so the previous guard silently treated ``{"cursor": true}`` as
cursor ``1``.
* Recovery now returns ``max(valid cursors) + 1``. Under adversarial
corruption "first int scanning in reverse" is not the same thing, and
only ``max`` keeps the recovered cursor strictly greater than every
legitimate cursor still on disk.
* Non-int cursors are logged exactly once per ``MemoryStore``. Silently
dropping corrupted entries hides the root cause (an external writer
to ``memory/history.jsonl``); rate-limiting keeps the log clean when
the same poisoned file is read every turn.
All 7 tests from the original fix pass unchanged; 3 new tests pin the
invariants above.
Made-with: Cursor
_next_cursor now checks isinstance(cursor, int) before arithmetic,
falling back to a reverse scan of all entries when the last entry's
cursor is corrupted. read_unprocessed_history skips entries with
non-int cursors instead of crashing on comparison.
Root cause: external callers (cron jobs, plugins) occasionally wrote
string cursors to history.jsonl, which blocked all subsequent
append_history calls with TypeError/ValueError.
Includes 7 regression tests covering string, float, null, and list
cursor types.
`append_history` previously used `strip_think(entry) or entry.rstrip()`
as a safety net, so if the entire entry was a template-token leak (e.g.
`<think>reasoning</think>` or `<channel|>` alone), the raw leaked text
was still persisted to history — later re-introducing the very content
`strip_think` was meant to scrub, via consolidation / replay.
Persist the cleaned content directly. When cleanup empties a non-empty
entry, log at debug and store an empty-content record (cursor continuity
preserved). Adds 3 regression tests in test_memory_store.py covering:
- Well-formed thinking blocks are stripped before persistence.
- Pure-leak entries persist as empty, not as raw text.
- Malformed prefix leaks (`<channel|>`) also persist as empty.
- Track last_summary in maybe_consolidate_by_tokens() to persist the summary
- Change return to break in the consolidation loop to allow summary persistence
- Save summary to session.metadata['_last_summary'] for consistency with AutoCompact._archive()
- Ensures compressed content remains visible to the model via prepare_session() injection
Fixes#3274
When chat_with_retry returns an error response (finish_reason='error')
instead of raising an exception, archive() previously treated the error
message as a valid summary and wrote it to history.jsonl, while the
original session data was already cleared by /new — causing irreversible
data loss.
Fix: check finish_reason after the LLM call and raise RuntimeError on
error responses, which naturally falls through to the existing raw_archive
fallback. This preserves the original messages in history.jsonl instead
of losing them.
Fixes#3244
Follow-up to #3212, fully backward compatible:
- Extract the 14-day staleness threshold as `_STALE_THRESHOLD_DAYS` module
constant and pass it into the Phase 1 prompt template as
`{{ stale_threshold_days }}`. The number lived in three places before
(code threshold, prompt instruction, docstring); now there is one.
- Add `DreamConfig.annotate_line_ages` (default True = current behavior)
and propagate it through `Dream.__init__` and the gateway wiring in
cli/commands.py. Gives users a knob to disable the feature without a
code patch if an LLM reacts poorly to the `← Nd` suffix.
- Harden `_annotate_with_ages` against dirty working trees: when HEAD
blob line count disagrees with the working-tree content length, skip
annotation entirely instead of assigning ages to the wrong lines. The
previous `i >= len(ages)` guard only handled one direction of the
mismatch.
- Inline-comment the `max_iterations` 10→15 bump with a pointer to
exp002 so future blame has context.
- Add 4 regression tests: end-to-end `← 30d` reaches prompt, 14/15
threshold boundary, `annotate_line_ages=False` bypasses git entirely
(verified via `assert_not_called`), length-mismatch defense, and
template-var rendering.
Made-with: Cursor
Three improvements to Dream's memory consolidation:
1. Per-line git-blame age annotations: MEMORY.md lines get `← Nd` suffixes
(N>14) from dulwich annotate. SOUL.md/USER.md excluded as permanent.
LLM uses content judgment, not just age, to decide what to prune.
2. Dedup-aware Phase 1 prompt: reframed as dual-task (extract facts +
deduplicate existing files) with explicit redundancy patterns to scan for.
Validated through 20 experiments (exp-002 prompt + max_iter=15 was best,
averaging -1643 chars/5.4% compression per run).
3. Phase 1 analysis as commit body: dream git commits now include the full
Phase 1 analysis for transparency via /dream-log.
4. max_iterations raised from 10 to 15: 30% improvement over 10 with no
risk; 20 showed diminishing returns (exp-020: -701 vs exp-017: -1643).
- Use .get('cursor') instead of direct dict access to prevent KeyError
- Skip entries without cursor and log a warning
- Fix _next_cursor fallback to safely check for cursor existence
Fixes#3190
Point Dream skill creation at a readable builtin skill-creator template, keep skill writes rooted at the workspace, and document the new skill discovery behavior in README.
Made-with: Cursor
Instead of a separate skill discovery system, extend Dream's two-phase
pipeline to also detect reusable behavioral patterns from conversation
history and generate SKILL.md files.
Phase 1 gains a [SKILL] output type for pattern detection.
Phase 2 gains write_file (scoped to skills/) and read access to builtin
skills, enabling it to check for duplicates and follow skill-creator's
format conventions before creating new skills.
Inspired by PR #3039 by @wanghesong2019.
Co-authored-by: wanghesong2019 <wanghesong2019@users.noreply.github.com>
Make Consolidator.archive() return the summary string directly instead
of writing to history.jsonl then reading back via get_last_history_entry().
This eliminates a race condition where concurrent _archive calls for
different sessions could read each other's summaries from the shared
history file (cross-user context leak in multi-user deployments).
Also removes Consolidator.get_last_history_entry() — no longer needed.
history.jsonl may contain non-UTF-8 bytes (e.g. from email channel
binary content), causing auto compact to fail when reading the last
entry for summary generation. Catch UnicodeDecodeError alongside
FileNotFoundError and JSONDecodeError.
When a user is idle for longer than a configured TTL, nanobot **proactively** compresses the session context into a summary. This reduces token cost and first-token latency when the user returns — instead of re-processing a long stale context with an expired KV cache, the model receives a compact summary and fresh input.
- Wrap both token estimation calls in try/except to prevent silent failures
from crashing the consolidation cycle
- Add _MAX_CHUNK_MESSAGES = 60 to cap messages per consolidation round,
avoiding oversized chunks being sent to the consolidation LLM
- Improve idle log to include unconsolidated message count for easier debugging
These are purely defensive improvements with no behaviour change for
normal sessions.
* feat(dream): enhance memory cleanup with staleness detection
- Phase 1: add [FILE-REMOVE] directive and staleness patterns (14-day
threshold, completed tasks, superseded info, resolved tracking)
- Phase 2: add explicit cleanup rules, file paths section, and deletion
guidance to prevent LLM path confusion
- Inject current date and file sizes into Phase 1 context for age-aware
analysis
- Add _dream_debug() helper for observability (dream-debug.log in workspace)
- Log Phase 1 analysis output and Phase 2 tool events for debugging
Tested with glm-5-turbo: MEMORY.md reduced from 149 to 108-129 lines
across two rounds, correctly identifying and removing weather data,
detailed incident info, completed research, and stale discussions.
* refactor(dream): replace _dream_debug file logger with loguru
Remove the custom _dream_debug() helper that wrote to dream-debug.log
and use the existing loguru logger instead. Phase 1 analysis is logged
at debug level, tool events at info level — consistent with the rest
of the codebase and no extra log file to manage.
* fix(dream): make stale scan independent of conversation history
Reframe Phase 1 from a single comparison task to two independent
tasks: history diff AND proactive stale scan. The LLM was skipping
stale content that wasn't referenced in conversation history (e.g.
old triage snapshots). Now explicitly requires scanning memory files
for staleness patterns on every run.
* fix(dream): correct old_text param name and truncate debug log
- Phase 2 prompt: old_string -> old_text to match EditFileTool interface
- Phase 1 debug log: truncate analysis to 500 chars to avoid oversized lines
* refactor(dream): streamline prompts by separating concerns
Phase 1 owns all staleness judgment logic; Phase 2 is pure execution
guidance. Remove duplicated cleanup rules from Phase 2 since Phase 1
already determines what to add/remove. Fix remaining old_string -> old_text.
Total prompt size reduced ~45% (870 -> 480 tokens).
* fix(dream): add FILE-REMOVE execution guidance to Phase 2 prompt
Phase 2 was only processing [FILE] additions and ignoring [FILE-REMOVE]
deletions after the cleanup rules were removed. Add explicit mapping:
[FILE] → add content, [FILE-REMOVE] → delete content.
Dream Phase 2 uses fail_on_tool_error=True, which terminates the entire
run on the first tool error (e.g. old_text not found in edit_file).
Normal agent runs default to False so the LLM can self-correct and retry.
Dream should behave the same way.
- Added Jinja2 template support for various agent responses, including identity, skills, and memory consolidation.
- Introduced new templates for evaluating notifications, handling subagent announcements, and managing platform policies.
- Updated the agent context and memory modules to utilize the new templating system for improved readability and maintainability.
- Added a new dependency on Jinja2 in pyproject.toml.
- Add GitStore class wrapping dulwich for memory file versioning
- Auto-commit memory changes during Dream consolidation
- Add /dream-log and /dream-restore commands for history browsing
- Pass tracked_files as constructor param, generate .gitignore dynamically
Add "Solutions" category to consolidate prompt so trial-and-error
workflows that reach a working approach are captured in history for
Dream to persist. Remove overly broad "debug steps" skip rule that
discarded these valuable findings.
Replace single-stage MemoryConsolidator with a two-stage architecture:
- Consolidator: lightweight token-budget triggered summarization,
appends to HISTORY.md with cursor-based tracking
- Dream: cron-scheduled two-phase processor that analyzes HISTORY.md
and updates SOUL.md, USER.md, MEMORY.md via AgentRunner with
edit_file tools for surgical, fault-tolerant updates
New files: MemoryStore (pure file I/O), Dream class, DreamConfig,
/dream and /dream-log commands. 89 tests covering all components.
Trigger token consolidation before prompt usage reaches the full context window so response tokens and tokenizer estimation drift still fit safely within the model budget.
Made-with: Cursor
Replace fire-and-forget consolidation with archive_messages(), which
retries until the raw-dump fallback triggers — making it effectively
infallible. /new now clears the session immediately and archives in
the background. Pending archive tasks are drained on shutdown via
close_mcp() so no data is lost on process exit.
- Require both history_entry and memory_update, reject null/empty values
- Fallback to tool_choice=auto when provider rejects forced function call
- After 3 consecutive consolidation failures, raw-archive messages to
HISTORY.md without LLM summarization to prevent context window overflow
Some providers (e.g. Dashscope in thinking mode) reject object-style
tool_choice with "does not support being set to required or object".
Retry once with tool_choice="auto" instead of failing silently.
Made-with: Cursor
Fix issue #1823: Memory consolidation does not inherit agent temperature
and maxTokens configuration.
The agent's configured generation parameters were not being passed through
to the memory consolidation call, causing it to fall back to default values.
This resulted in the consolidation response being truncated before the
save_memory tool call was emitted.
- Pass temperature, max_tokens, reasoning_effort from AgentLoop to
MemoryConsolidator and then to MemoryStore.consolidate()
- Forward these parameters to the provider.chat_with_retry() call
Fixes#1823
Move consolidation policy into MemoryConsolidator, keep backward compatibility for legacy config, and compress history by token budget instead of message count.
Major changes:
- Replace message-count-based memory window with token-budget-based compression
- Add max_tokens_input, compression_start_ratio, compression_target_ratio config
- Implement _maybe_compress_history() that triggers based on prompt token usage
- Use _build_compressed_history_view() to provide compressed history to LLM
- Refactor MemoryStore.consolidate() -> consolidate_chunk() for chunk-based compression
- Remove last_consolidated from Session, use _compressed_until metadata instead
- Add background compression scheduling to avoid blocking message processing
Key improvements:
- Compression now based on actual token usage, not arbitrary message counts
- Better handling of long conversations with large context windows
- Non-destructive compression: old messages remain in session, but excluded from prompt
- Automatic compression when history exceeds configured token thresholds
