Scope Recall for Hermes

Hermes current-turn memory provider with journal-first semantic capture, durable recall, SQLite truth storage, and optional vector companions

Give Hermes durable memory that can follow the same user across windows/chats while keeping local scratch context from bleeding into the wrong place.

Current-turn recall · Journal-first capture · Durable shared memory · Background digest · Local scratch scopes · SQLite truth · LanceDB/SQLite companion · Hybrid RRF retrieval

scope-recall is a Hermes local memory provider built for current-turn recall and durable semantic memory. Durable user/project/ops/memory facts are shared across windows/chats for the same user + agent identity; raw general turn captures stay local to the current chat/thread/session.

This repository, scope-recall-hermes, is the Hermes implementation. The Python distribution package is hermes-scope-recall, the Python import/package spelling is scope_recall, and the Hermes plugin ID/provider name remains scope-recall for runtime compatibility. The OpenClaw sibling implementation lives at scope-recall-openclaw.

Benchmark: LoCoMo long-conversation memory

In August 2026 we ran Scope Recall's 1.9.2 development line end-to-end on the LoCoMo long-conversation benchmark: all 1,540 questions across the four non-adversarial categories, with 100% coverage and zero invalid results — a full run, not a sample.

• Overall accuracy: 70.58% (1,087 / 1,540)

• Single-hop factual memory: 84.78%

• Temporal reasoning: 54.83%

• Multi-hop reasoning: 54.61%

• Open-domain inference: 45.83%

• Retrieval side: Top-50 evidence any-hit recall 97.66%, all-hit recall 90.63%

The run used gemini-embedding-001 for retrieval and gpt-5.4-mini as answerer, judge, and query planner. Every per-question result and the run manifest are hash-pinned, and the final score was recomputed independently from the raw rows; the public, path-free receipt is docs/benchmarks/locomo-2026-08.md. The benchmark harness (scripts/benchmark.locomo.py) is part of the 1.9.2 development line so the evaluation stays reproducible.

The August run predates the harness's secret-free model-route receipt and stricter official-comparability checklist. It is therefore legacy local evidence, not a run for which the current official_comparable_categories_1_to_4 flag may be asserted.

The runner has no machine-specific dataset, source-tree, or auth defaults. Supply the external paths explicitly, keep the run directory outside the source checkout, and add --auth-path only for model-backed phases:

python scripts/benchmark.locomo.py \
  --dataset /path/to/locomo.json \
  --run-dir /path/outside/the/checkout/locomo-run \
  --hermes-agent-root /path/to/hermes-agent \
  --phase retrieve

The three categories near 50% remain the clearest future improvement areas: multi-hop evidence completeness, temporal evidence presentation, and open-domain synthesis. We do not claim a cross-vendor ranking, because public LoCoMo evaluations use materially different models, judges, prompts, and dataset variants.

Version 1.10.3 is the public patch candidate on the last packaged 1.10.2 line. It fixes issue #50: governance coverage and cleanup rollback now recognize the official memory_auto_adjudication + archive receipt, while unknown archive writers remain fail-closed. Rollback still refuses state changed after the recorded receipt. SQLite authority, stable provider/tool identities, schema, and automatic-adjudication policy are unchanged.

The 1.8.6 patch hardened legacy fact-freshness maintenance, Unicode secret filtering, operator portability, truth-connection ownership, lifecycle relation restore, and semantic deduplication. It preserves the 1.8.5 Windows activation-lease PID fix and includes the 1.8.4 deep-audit closure for deterministic freshness, validator, truth-store permission, operator recovery, secret scanning, and cross-platform release gates. Its 1.8.3 predecessor hardened current-state recall, canonical chat identity boundaries, Windows repair and rollback, bounded vector-outbox history, and model-calibrated vector-only filtering. The 1.8.2 release made memory startup honest and local-model tool use reliable. If a configured local embedding model cannot actually load, Scope Recall now reports the degradation and uses a compatible fallback only for a fresh generation; it never opens an existing generation with a different embedding space. It also adds light, balanced, and full semantic retention profiles and fixes LM Studio/llama.cpp grammar initialization without deleting structured claim, freshness, or evolution capabilities. Durable user, memory, project, and ops fact actions resolve to shared durable scope, while general remains local scratch on every integration path. Existing ordinary-memory behavior remains the default because the new evolution, temporal-query, and Reflection surfaces are opt-in. The 1.7.2 release published a compatibility-preserving storage and governance hardening patch: ordinary recall uses one lifecycle policy across journal, nightly, deduplication, and vector paths; vector rebuilds support immutable generations and explicit compare-and-swap activation; metadata and import provenance are sanitized before durable or operator-visible sinks; candidate, freshness, and config mutations fail closed; and folded inline data URLs are removed without losing surrounding prose. It builds on version 1.7.1's runtime-config, candidate-browser, external-bridge, and release-gate fixes. Version 1.7.0 published the productization feature set on the stable V1 release line: event-digest evidence packets, reviewable candidate extraction, read-only memory browsing, candidate governance commands, Experience-to-skill bridge helpers, optional PGVector companion support, external shared-memory bridge contracts, explicit sensitivity governance, release-gate progress output, and same-process peer-provider SQLite lock recovery for scope_recall_store. Version 1.6.3 closed issue #25 with conservative SQLite lock recovery and a single safe retry for scope_recall_store while keeping non-SQLite business errors non-retryable. Version 1.6.2 added graph-relation backfill/benchmark visibility and hardened Experience review and journal-digest bookkeeping without changing the stable V1 runtime contract. Version 1.6.1 published documentation, packaging, and release-provenance updates without changing the stable V1 runtime contract. The 1.6.0 release packages a compatibility-preserving refactor of the doctor, graph-hygiene, maintenance, digest-result, recall-pipeline, and provider-schema internals while keeping the stable V1 commercial-governance line introduced in 1.5.0. The 1.5 line includes promoted-only profile lifecycle safety, candidate-memory promotion planning, graph-hygiene repair, fail-closed vector-repair fallback handling, governance cleanup, journal recovery, an operator dashboard, repository-owned golden benchmarks, stricter release gates, fail-closed hard-delete safety, packaged benchmark fixtures, Recall Funnel observability, synthetic retrieval-regression benchmarking, and default-safe vector fallback behavior. Runtime Experience packet injection is enabled by default through experience.prefetch_enabled=true and can be disabled with experience.prefetch_enabled=false; background automatic promotion remains an explicit operator opt-in through experience.auto_promotion_enabled=true, and low-risk auto-promotion remains a second explicit opt-in through experience.auto_promote_low_risk=true. By default, successful low-risk scans create candidate playbooks, high-risk playbooks stay review-gated, and final-failure or low-signal traces are not promoted. It keeps the scope_recall_profile surface added in v1.3.0, compression-boundary journal staging through Hermes' on_pre_compress() memory-provider hook, inline attachment-marker sanitization, the supported standalone install shape added in v1.1.0, and native-safe LanceDB probing with fresh-bootstrap SQLite vector fallback for non-AVX hosts.

It uses a three-layer design:

• Journal/provenance layer for eligible raw conversation turns and evidence links that are not directly recalled as durable memory

• SQLite truth store for high-density durable local records and deterministic auditing

• Vector companion for semantic retrieval and hybrid ranking: LanceDB by default, or sqlite-bruteforce for native-free/non-AVX hosts

This replaces the old lancepro naming, which was misleading because the earlier implementation was SQLite-only.

Design promises

• Truth stays inspectable: SQLite remains the authoritative store; vectors are rebuildable.

• Recall is current-turn scoped: retrieval is based on the active query, not stale queued context from the previous topic.

• Durable memory travels deliberately: user, memory, project, and ops facts can follow the same user + agent identity across windows/chats, and can cross platforms only when explicit canonical identity mapping is configured.

• Raw turns are provenance, not durable memory: eligible conversation turns are written to a journal/staging layer; only