feat: Mem0-style two-phase extraction+update pipeline#2
Merged
Conversation
Implements FactExtractor class and ExtractedFact dataclass using ollama qwen2.5:3b to distill raw content into importance-scored facts (1-10 scale), with markdown fence stripping, JSON parsing, descending sort, and graceful error fallback. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Wires FactExtractor (Phase 1) and MemoryUpdater (Phase 2) together via a new remember_with_extraction() method on MemoryManager, enabling Mem0-style ADD/UPDATE/DELETE/NOOP decisions per extracted fact. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Remove dead min_importance param from FactExtractor (filtering done in MemoryManager) - Harden importance parsing to handle non-numeric LLM output - Rename DELETE status from "deleted" to "corrected" (soft-delete, not actual deletion) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
5 tasks
Owner
Author
|
LGTM. Clean implementation, tests pass. Next step: wire remember_with_extraction() into the LOCOMO benchmark harness and re-run to measure improvement. |
rolandpg
added a commit
that referenced
this pull request
Apr 15, 2026
…mark design Performance Benchmarker analysis covering: - Survey of 20+ benchmarks across 7 categories - Priority matrix: CTIBench ATE fix (#1), CTI-specific v2 (#2), scale 10K (#3), evolve quality (#4), RAGAS CTI (#5) - Custom CTI benchmark design with alias-aware scoring - Harness architecture for CI-integrated regression detection - Bug found: auto_ralph.py numpy import at wrong scope Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
rolandpg
added a commit
that referenced
this pull request
Apr 15, 2026
Data Engineer analysis of live OpenCTI (114K rels, 140K indicators, 73K observables) vs ZettelForge data model reveals ZettelForge cannot represent the dominant IOC workload. Key gaps: no indicator type, no observable types (hashes/IPs/domains), no ATT&CK pattern entity, no TLP markings, no STIX pattern parser. 6 prioritized fixes identified. IOC regex extractors (#1) and AttackPattern entity (#2) are highest impact at lowest effort. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
3 tasks
rolandpg
added a commit
that referenced
this pull request
Apr 26, 2026
Three blockers found in code review of the RFC-015 web GUI:
1. config_page (/config HTML route) called _to_dict, which was defined as
a nested function inside get_config_endpoint. Every render hit a
NameError that was silently swallowed by a bare except, leaving
config_yaml = "" and forcing the SPA to fall back to /api/config.
Promoted _to_dict to a module-level _config_to_dict helper used by
both routes.
2. update_config compared top-level payload keys against a set of
dotted-path restart-required fields, so {"embedding": {"provider": "x"}}
was reported as applied: ["embedding"], pending_restart: [], silently
telling operators a restart-required change had taken effect.
Added _flatten_keys to walk the nested payload to dotted leaf paths
and lifted the restart set to module scope as _RESTART_REQUIRED_FIELDS.
3. /config HTML route had no auth guard. /api/config was protected, but
the page shell (and once #1 was fixed, the server-rendered YAML body)
was reachable without an API key. Added Depends(require_api_guard)
to the route and made the YAML body redact secrets before serializing.
Tests:
Added four regression tests in tests/test_web_api.py covering nested
restart-required flags, nested non-restart application, /config auth
gate (via monkeypatch of API_KEY because it is captured at module import
time), and /config server-side YAML render. 24 passed, 2 skipped.
Note on the original review's blocker #2 (multi-tenancy regression on
graph and entity endpoints): not a regression. There is no per-tenant
MemoryManager._knowledge_graph anywhere in community or enterprise.
get_knowledge_graph() is a process-wide singleton; get_mm_for_request
is a stub returning the default MemoryManager. The old code's
getattr(tenant_mm, "_knowledge_graph", None) always returned None,
so the new singleton call is a functional improvement, not a regression.
No code change needed.
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
remember_with_extraction()method on MemoryManager orchestrates both phasesWhy
ZettelForge scored 14% on LOCOMO benchmark vs. Mem0's 66.9% SOTA. Root cause: append-only storage creates redundancy, contradictions, and noise. This pipeline adds selective extraction (only important facts stored) and coherence operations (UPDATE supersedes stale info, DELETE corrects contradictions, NOOP skips duplicates).
New files
src/zettelforge/fact_extractor.py— Phase 1 extractionsrc/zettelforge/memory_updater.py— Phase 2 operationstests/test_fact_extractor.py— 11 unit teststests/test_memory_updater.py— 14 unit teststests/test_two_phase_e2e.py— 5 integration testsTest plan
🤖 Generated with Claude Code