yadgar 5.141.0

Persistent memory engine for Claude Code — heat decay, sleep consolidation, and surprise-gated storage

Changelog · Benchmark · Architecture · Roadmap · JS/TS SDK · Agents guide

AI coding agents: the operational guide lives in AGENTS.md — setup commands, dev environment, test runner, code style, PR rules, security gates. This README is the human overview.

A persistent memory engine for Claude Code. Tell it what matters and it survives across sessions — decaying what you stop touching, promoting what repeats, filtering recall to the git branch you're on, and pairing every memory with a curated wiki that searches through the same ranking pipeline.

Yadgar (یادگار) is Persian for "memento, keepsake."

---

What it is

Claude Code forgets everything when a session ends, when you /clear, or when context compacts. Yadgar is the layer that remembers. It runs as an MCP server alongside Claude Code and gives it two durable stores:

1. Memory — episodic + semantic facts, each carrying a heat value that decays over time. Access reheats; disuse decays; recurring episodes get promoted to semantic memory by a nightly consolidation cycle modelled on how brains sleep.
2. A curated wiki — long-lived, versioned, branch-scoped pages: conventions, module purpose, past decisions, where subsystems live. Written deliberately, not captured incidentally.

A single recall() query searches both stores at once, fuses and re-ranks the results, weights them by heat, gates out decayed entries, and boosts matches on your current git branch. Critical context can be anchored so it never decays; whole working states can be checkpointed and restored across a /compact.

Why

• Survives sessions. Heat decay drops what you stopped using; surprise gating drops duplicates on arrival; anchors pin what must never be lost.
• Branch-aware. Recall boosts current-branch matches 1.5×; wiki pages resolve in branch precedence (current → default → unscoped) so canonical knowledge stays reachable from feature work.
• One ranking pipeline for memory + wiki. Curated knowledge and episodic memory come back in a single ranked result, not two separate searches.
• Consolidates while you sleep. A nightly brain cycle decays heat, promotes episodes to semantics, discovers causal links, forms associative links, and merges duplicates — without dropping the MCP connection.
• Self-documenting. Architecture Decision Records, a reusable agent-prompt library, and an interactive 3D knowledge-graph all live in the same store.

---

Key features

Memory

• Heat-decay lifecycle — every memory carries heat ∈ [0,1] decaying exponentially; access reheats, disuse decays, unused items eventually archive and purge.
• Surprise-gated writes — a similarity gate rejects near-duplicate writes on arrival (rejections land in the DLQ, inspectable via dlq_inspect).
• Write-time contradiction detection — a lightweight negation + action-divergence heuristic flags contradicting memories and decays their confidence, without blocking the write.
• Anchors — protected memories that never decay; audit_anchors() surfaces redundant, oversize, expired, or near-duplicate (cosine ≥ 0.95) anchors across projects.
• Checkpoint / restore — checkpoint(directory) snapshots your working state before compaction; restore(directory) reconstructs it afterward via a hippocampal-replay blend of checkpoint + anchors + hot memories + predicted context. Use restore, not recall, to resume after /clear or /compact.
• Memory blocks — Letta-style named, scoped, char-limited core-memory containers (block_*) for always-in-context working state.

Curated wiki

• Versioned pages — full history (wiki_history, wiki_diff, wiki_read_version, wiki_restore); wiki_add commits directly (no draft step).
• Branch-scoped resolution — wiki_read(slug) resolves current branch → default branch → unscoped.
• Surgical editing — anchor-text and positional edit tools (wiki_replace_text, wiki_insert_before/after, wiki_append_section, wiki_replace_markdown_block, positional wiki_*_at) so pages mutate in place instead of full rewrites.
• Auto-linking — wiki_autolink inserts [[slug]] cross-references by matching page titles; validated so it never manufactures broken refs.
• Repo wiki — repo_wiki_generate builds code-structure pages from a Python repo via AST scanning (no LLM), writing directly into the wiki store with SHA256-parity staleness detection. Sync helpers: wiki_refresh_stale, wiki_cleanup_merged_branches, wiki_coverage, wiki_lint.
• Bookmarks — pin wiki pages in the viz UI (bookmark_*); drag-to-reorder, dense-integer positions.

Unified recall

• One recall() query retrieves memory + wiki together, fused through WRRF, cross-encoder reranked, NLI- and MMR-filtered, and run through the rules engine — heat-weighted, decay-gated, branch-scoped. Filter by type= ("memory", "wiki", or both). wiki_query remains for wiki-only search.

Nightly consolidation (the "brain cycle")

Runs nightly while the daemon stays up in maintenance mode (no MCP reconnect). Phases: apply_decay → process_episodes → merge_duplicates → link_similar → detect_causality → memify → cls_consolidation plus a dream/sleep phase.

• CLS promotion — recurring episodic patterns promote to semantic memory (Complementary Learning Systems).
• Causal discovery — PC-algorithm causal-DAG inference over co-occurring memories.
• Dream replay — associative co_occurrence linking + insight generation; insights cap at 21 days if unaccessed.
• Community detection — graph clustering surfaced as clusters in the viz.
• Re-embeds stale memories; optionally precomputes the graph layout (below).

Architecture Decision Records

• adr_add writes to a per-project ADR log in the wiki store, enforcing an 11-field schema and auto-incrementing IDs (ADR-0001, ADR-0002, …). A Stop-hook prompt captures decisions at session end. Yadgar dogfoods its own ADR system (ADR-0004, 0007, 0011, 0013 are all real entries).

Agent-prompt library

• Reusable subagent dispatch prompts, stored as tagged wiki pages (agent-prompt-<pattern>), versioned and improving over time. agent_prompt_save(pattern, content, purpose) upserts one page per pattern; agent_dispatch_prelude(pattern) builds a prelude to prepend to a subagent prompt; seed_agent_prompts() seeds four starter patterns. Lookup is via tagged recall (recall(type="wiki", tags=["agent-prompt"])), kept out of normal recall noise.

Knowledge-graph viz

• yadgar viz serves an interactive force-directed graph of memories, wiki pages, entities, and relationships at http://localhost:42069 (2D + 3D; memory = sphere, wiki = octahedron, anchor = cube; heat encoded as color). Filter by node type, edge type, and heat; focus mode, hover-neighborhood highlight, search, and connection-count badges.
• Precomputed server-side layout (VIZ_PRECOMPUTED_LAYOUT_ENABLED, default off): the nightly cycle computes 3D node positions once (networkx.spring_layout, signature-cached) so the graph renders pre-laid-out instead of a cold client-side layout per load. Otherwise the client warm-starts from localStorage.
• The UI is organized into four menus: Graph · Bookmarks · System {Config, Health, Stats} · Help {Guide, Config Reference, About, Debug}.

Config system

• pydantic-settings with precedence env vars (YADGAR_*) > ~/.config/yadgar/config.yaml > defaults.
• System → Config is an in-browser editor (part of the viz UI): category-grouped knobs, alpha-sorted, hover tooltips, ⓘ deep-links to a Config Reference page, and SOURCE badges distinguishing env-var / config-file / default origins. Writes are bearer-auth gated.

Security & observability

• Bearer-token MCP auth on /api/*, /hooks/*, /mcp — default-deny CORS, timing-safe compare, loopback-only by default. /health and /metrics exempt on loopback.
• Always-on secret gate blocks AWS / GCP / Stripe / Slack / OpenAI / Anthropic keys, JWTs, GitHub PATs, private keys, and DB URIs before they reach the store (cannot be disabled; context-aware allowlist for known-good fixtures).
• Auto-capture sanitization strips ANSI escapes, control chars, and Unicode bidi-override before action-log insert.
• Prometheus /metrics + OpenTelemetry distributed tracing across core + backend (core→backend traceparent joins one trace); structured JSON logs; per-phase consolidation duration markers. Tri-signal standard (v5.101, ADR-0034): every in-scope function emits span+metric+log via the @observe decorator, ratcheted by the I33 coverage lint.
• Async write queue with retry/backoff, dead-letter for permanent failures, and DLQ inspection tools (dlq_inspect, dlq_requeue, dlq_dismiss).

---

Architecture

┌──────────────┐         MCP (stdio / streamable-http)        ┌──────────────────────────┐
│  Claude Code │ ◄──────────────────────────────────────────► │  yadgar core (:8765)     │
│   + hooks    │   memorize / recall / wiki_* / checkpoint …   │  FastAPI + MCP server    │
└──────────────┘                                               │  retrieval pipeline      │
                                                               │  consolidation scheduler │
                                                               │  viz web UI (:42069)     │
                                                               └────────────┬─────────────┘
                                                                            │ HTTP
                                                               ┌────────────▼─────────────┐
                                                               │  yadgar-backend           │
                                                               │  SurrealDB store (:8000)  │
                                                               │  embed + rerank (:8001)   │
                                                               └───────────────────────────┘

• yadgar core — the MCP server (FastAPI + the mcp SDK). Hosts the retrieval pipeline, the nightly consolidation scheduler, the rules engine, the security middleware, and the viz web UI on :42069. Talks to Claude Code over stdio or streamable-HTTP.
• yadgar-backend — SurrealDB as the vector + full-text store (:8000) and the embedding / cross-encoder rerank service (:8001). Versioned on an independent track from core.
• Retrieval pipeline — recall() runs FTS + KNN vector + personalized-PageRank + spreading-activation + temporal candidate generation → WRRF fusion → cross-encoder rerank → NLI → MMR diversity → adversarial detection → rules engine. Branch filter (branch IN (current, default, NULL)) applies post-fetch; current-branch matches get a 1.5× boost. Behavior is pinned by characterization tests.
• Consolidation — a nightly cycle (19:00 UTC) plus a weekly vacuum; runs in-process while the daemon stays up.

Deeper detail: docs/reference/architecture.md · docs/reference/retrieval.md · docs/reference/memory-lifecycle.md.

Layer docs (in-tree): each layer root carries a README.md (subsystem map) and an AGENTS.md (placement laws for coding agents) — yadgar/_shared/ · yadgar/backend/ · yadgar/core/. Major subsystem packages (storage, retrieval, config, observability, security, wiki, embed_service, consolidation, server, viz, daemon, cli, install, seed, hooks) carry their own README.md; existence is lint-enforced (scripts/check_subsystem_readmes.py).

---

Install

Python 3.14+ on the host (or use the Docker path for zero host Python). All paths reach the same yadgar setup post-install step.

pipx (recommended for isolated install):

pipx install yadgar
yadgar setup

Nix flake:

nix profile install codeberg:maxagahi/yadgar
yadgar setup

Plain pip:

pip install yadgar
yadgar setup

Repo checkout (development):

git clone https://codeberg.org/maxagahi/yadgar.git
cd yadgar
make setup            # install + hooks + agents + units + seed anchors

yadgar setup writes ~/.config/yadgar/config.yaml, generates ~/.config/yadgar/secrets.env (chmod 600) with random YADGAR_MCP_AUTH_TOKEN + SURREAL_PASS + YADGAR_RW_PASS + YADGAR_RO_PASS, installs Claude Code hooks + subagent templates + rules, seeds anchors, and prepares systemd (Linux) or launchd (macOS) user units. Idempotent — re-run after upgrades.

Then start the daemon and register it with Claude Code:

set -a && . ~/.config/yadgar/secrets.env && set +a
yadgar daemon start
yadgar daemon configure-mcp   # writes ~/.claude.json with Authorization: Bearer header

Restart Claude Code, then verify:

yadgar daemon status
yadgar stats

Per-platform detail: docs/reference/install.md.

Stdio-only (no daemon, no Docker)

Single-session use. Skip yadgar setup. Add to ~/.claude.json:

{ "mcpServers": { "yadgar": { "command": "yadgar", "args": [] } } }

Restart Claude Code. No bearer auth; embed / rerank degrade gracefully without the backend.

Docker

Two containers — backend (SurrealDB + embed service) and core (MCP server + viz). Note the independent version tracks: core tracks the project version, backend tracks its own.

docker network create yadgar-net

docker run -d --name yadgar-backend --network yadgar-net \
  -v yadgar-db-data:/data \
  -e SURREAL_USER=$SURREAL_USER \
  -e SURREAL_PASS=$SURREAL_PASS \
  openfantasy/yadgar-backend:5.43.0

docker run -d --name yadgar --network yadgar-net \
  -v yadgar-data:/data \
  -p 127.0.0.1:8765:8765 \
  -p 127.0.0.1:42069:42069 \
  -e YADGAR_DB_URL=http://yadgar-backend:8000 \
  -e YADGAR_EMBED_URL=http://yadgar-backend:8001 \
  -e YADGAR_MCP_AUTH_TOKEN=$YADGAR_MCP_AUTH_TOKEN \
  openfantasy/yadgar:5.132.0

Containers bundle Python 3.14 — no host Python required. Source ~/.config/yadgar/secrets.env (or generate the required vars yourself) before launching.

Updating

yadgar update --check   # probes PyPI; prints upgrade command; exit 0

For pipx the suggested command is pipx upgrade yadgar; then re-run yadgar setup (idempotent) to refresh hooks and units. An opt-in update orchestrator (yadgar update --install, default off via update.install_enabled: false) coordinates snapshot → image pull → graceful drain → restart → health-check → CLI upgrade, with automatic rollback (yadgar update --rollback). See MIGRATION_NOTES.md.

---

CLI

yadgar                              # MCP server (stdio); --transport {stdio,sse,streamable-http}
yadgar daemon start|stop|logs|health|status
yadgar daemon configure-mcp         # write ~/.claude.json with bearer header
yadgar setup                        # first-run config + secrets + MCP snippet
yadgar stats [--project /path]      # memory statistics
yadgar viz                          # knowledge graph at http://localhost:42069
yadgar vacuum                       # compact the SurrealKV store
yadgar seed <directory>             # bootstrap memory for an existing project
yadgar repo-wiki <directory>        # generate code-structure wiki pages
yadgar rules export|import          # policy rules
yadgar config init|list|get|set     # configuration
yadgar update --check|--install|--rollback
yadgar export duckdb --output snap.duckdb   # analytics snapshot (pip install yadgar[analytics])
yadgar install-hooks|install-subagents      # wire Claude Code hooks + agent templates

---

MCP tools

Yadgar exposes 75 MCP tools. ⚡ marks power=True tools (gated in the minimal MCP profile). A categorized selection — full list via the running server's tool catalog:

Memory

Tool	Power	Purpose
memorize(content, context, tags)		Store a memory; auto-captures branch + surprise gate
recall(query, type, max_results)		Unified branch-aware search over memory and wiki
recent_memories()		Newest-first listing, no classifier dependency
memory_get(id)		Fetch by numeric ID
memory_update(id, fields)	⚡	Patch content / tags / flags
forget(id)		Mark for deletion
anchor(content, context, reason)	⚡	Protected memory; never decays
checkpoint(directory, …)	⚡	Snapshot pre-compaction
restore(directory)	⚡	Reconstruct post-compaction (hippocampal replay)
memory_stats()		Health + counts
validate_memory()	⚡	Check validity against current file state

Wiki (core + editing + maintenance)

Core: wiki_add · wiki_read ⚡ · wiki_query · wiki_list ⚡ · wiki_get · wiki_update ⚡ · wiki_delete ⚡ · wiki_history · wiki_diff · wiki_read_version · wiki_restore ⚡ · wiki_check_duplicate · wiki_lint ⚡

Editing: wiki_append_section ⚡ · wiki_replace_text ⚡ · wiki_delete_text ⚡ · wiki_insert_before/after ⚡ · wiki_replace_at / wiki_delete_at / wiki_insert_at ⚡ · wiki_replace_markdown_block ⚡ · wiki_autolink ⚡ · wiki_set_metadata ⚡

Maintenance: wiki_coverage · wiki_refresh_stale ⚡ · wiki_cleanup_merged_branches ⚡ · repo_wiki_generate

wiki_add commits directly — there is no draft/approve workflow on the production write path. (wiki_drafts / wiki_approve / wiki_discard exist but no path produces drafts.)

Bookmarks & Blocks

Bookmarks: bookmark_add · bookmark_remove · bookmark_list · bookmark_reorder

Blocks (Letta-style core memory, all ⚡): block_create · block_get · block_update · block_delete · block_list · block_replace · block_append

Project state

Tool	Power	Purpose
project_brief(directory, mode)		Layered bootstrap — catalog (~500 tok), full (~1050 tok), or signals
bootstrap_project(directory, content)	⚡	Set _project_init
update_active_work(directory, content)	⚡	Atomic replace of _active_work
seed_project(directory)	⚡	Bootstrap memory from README + top-level docs
install_hooks(scope)	⚡	Wire Claude Code hooks; inject bearer token

get_project_context() is a deprecated alias of project_brief(mode="catalog").

Consolidation & ops

consolidate_now() ⚡ · vacuum_now() ⚡ · vacuum_checkpoints() ⚡ · reembed_all() ⚡ · check_invariants(repair) ⚡ · sync_instructions() ⚡ · archive_purge() ⚡ · add_rule(...) ⚡ · get_rules(...) ⚡ · audit_anchors() ⚡

Agent-prompt library & ADR

Agent prompts: agent_prompt_save(pattern, content, purpose) ⚡ · agent_dispatch_prelude(pattern) · seed_agent_prompts() ⚡

ADR: adr_add(...) ⚡ — append an 11-field Architecture Decision Record to the project ADR log

Dead-letter queue

dlq_inspect(filter) · dlq_requeue(id) ⚡ · dlq_dismiss(id) ⚡

A typed JavaScript / TypeScript SDK wraps the tool surface as async methods (Node.js, Vercel Edge, Cloudflare Workers, Deno) — see docs/reference/sdk-js.md.

---

Configuration

yadgar config init        # write ~/.config/yadgar/config.yaml
yadgar config set retrieval_profile fast

Priority: env vars (YADGAR_*) > ~/.config/yadgar/config.yaml > defaults. Knobs are also editable in-browser via System → Config in the viz UI. Key vars (full reference in docs/reference/configuration.md):

Var	Default	Purpose
YADGAR_REQUIRE_AUTH	1	Bearer auth on /api/* /hooks/* /mcp. Set 0 only during initial rollout.
YADGAR_MCP_AUTH_TOKEN	(required)	Bearer token. yadgar setup generates one.
YADGAR_DB_PASS	(required)	SurrealDB password. No root:root fallback.
YADGAR_HOST	127.0.0.1	Bind interface. Loopback by default.
YADGAR_ALLOWED_ORIGINS	loopback	CORS allowlist.
YADGAR_METRICS_ENABLED	1	Expose Prometheus /metrics (loopback, unauthenticated).
YADGAR_LOG_FORMAT	human	Set json for structured logs.
YADGAR_VIZ_MAX_MEMORIES / _WIKI / _ENTITIES	500 / 200 / 2000	Viz node caps (0/-1 = unlimited).
VIZ_PRECOMPUTED_LAYOUT_ENABLED	false	Serve nightly-precomputed 3D layout.
YADGAR_MODEL_IDLE_EVICTION_SECONDS	0	Unload heavy ML models after idle seconds (0 = stay loaded).

---

Subagent integration

Yadgar ships a SubagentStop hook that captures memory findings from Claude Code subagents: when a subagent completes, its final report is scanned for a ## Yadgar findings section and each bullet is persisted with provenance_agent set to the agent type.

To opt your subagents into the protocol, paste docs/reference/claude-subagent-contract.md into your ~/.claude/CLAUDE.md, then run yadgar install-hooks --scope global. The contract is opt-in — Yadgar works without it.

---

Benchmark

Yadgar is evaluated on LongMemEval (ICLR 2025), the standard academic benchmark for long-term conversational memory. The longmemeval_s variant runs 500 questions across 6 categories against ~50 sessions of synthetic history per query, scored on Phase 1 retrieval (does the memory layer surface gold-context sessions in top-k?) and Phase 2 QA accuracy (does the reader produce the gold answer, judged by an LLM grader?).

Headline (v5.26.0, full 500q, Sonnet 4.6 reader + judge — the most recent full run; subsequent versions targeted infrastructure, viz, wiki, and recall-unification rather than retrieval quality):

System	Reader	LongMemEval-s QA	vs yadgar
yadgar v5.26.0	Sonnet 4.6	69.4% (347/500)	—
mem0 V3	GPT-4o	94.4%	+25.0 pp
Zep / Graphiti	GPT-4o	63.8%	−5.6 pp

Yadgar beats Zep by 5.6 pp on the same 500-question sample. mem0 leads by 25 pp via LLM-extract-on-ingest. Phase 1 retrieval: MRR 0.928, Recall@10 0.906, NDCG@10 0.863 — the memory layer surfaces gold context for ~91% of queries, so the remaining QA gap is mostly reader synthesis, not retrieval.

Full methodology and per-type breakdown: docs/benchmark-results/BENCHMARK_RESULTS.md.

---

Roadmap

Shipped highlights (v5.78 → v5.104)

• Unified recall (v5.78–v5.81) — memory + wiki merged into one ranked, heat-weighted, branch-scoped result; now the default.
• ADR tooling (adr_add, v5.85) + capture-first Stop-hook prompt.
• Agent-prompt library rework (v5.85, ADR-0007) — wiki-backed, tagged-recall lookup.
• Wiki autolink + repo-wiki store-bridge (v5.85).
• Viz overhaul + precomputed server-side layout + System → Config editor (v5.86–v5.88).
• Recall perf + accounting (v5.97–v5.104) — fusion/MMR N+1 batches, GTE-ModernBERT reranker (Lever-1, v5.98), spreading-activation N+1 batched (v5.104); latency FULLY accounted via MCP-tool traces (fusion pass quality-load-bearing — ADR-0035). CE metric corrected post-T4: CE is ~25% of the cold recall wall, not the "~90%" once quoted here (dead-metric artifact — ADR-0105, #192).
• Ettin-32m CE reranker (Train 4) — CE primary swapped GTE-ModernBERT → cross-encoder/ettin-reranker-32m-v1 (~4.7× faster per-pass, 2.44× end-to-end recall; GTE-ModernBERT kept as config-revert rollback). ADR-0106: backend standing config --cpus 3.
• Tri-signal observability standard (v5.100–v5.101, ADR-0034) — span+metric+log per function via the @observe decorator, I33 coverage lint, core→backend traceparent, OTLP → Tempo.
• Test-speed train (v5.104, ADR-0036) — module-scoped storage fixture + batched SurrealDB wipe → CI shards ~2× faster.
• Observability train (v5.83) — /health 503-on-degraded, OTLP circuit breaker, off-loop span logs.

Major upcoming

• v6 — Nightly LLM curator. A local agent (Ollama; two-tier deepseek-r1 + qwen3:8b) runs each night to detect staleness, annotate contradictions, find semantic correlations beyond co-occurrence, propose merges/forgets, and dedupe wiki pages — skipping cleanly if Ollama is offline.
• v7 — Real-time synthesis. recall(synthesize=True) / wiki_query(synthesize=True) append a synthesized answer alongside raw records; a new ask() tool returns synthesis-only output. Depends on a sub-10 s local synthesis model.

Full history: CHANGELOG.md.

---

Documentation

• AGENTS.md — operational guide for AI coding agents (setup, dev env, tests, code style, PR rules)
• Architecture — component map, retrieval, security, observability
• Memory lifecycle — heat, archiving, pruning, consolidation phases
• Retrieval — fusion, rerank, branch filter, pipeline stages
• Configuration — configuration reference (env vars, precedence, key knobs)
• Install — per-platform setup + the yadgar setup --doctor probe
• JS/TS SDK — typed client for the MCP tool surface
• Release runbook · Migration notes · Subagent contract

---

Contributing

Every change to yadgar/** must update README.md and docs/ in the same PR. Conventional Commits format. No Co-Authored-By: trailers. AI coding agents working on this repo: read AGENTS.md for the full dev / test / PR contract.

Related projects

• ccpm — Claude Code Plugin Marketplace. Ships code-review, confluence-rfc, git-flow, repo-wiki, tf-naming-check, and update-jira plugins that compose with yadgar.

Tribute

Inspired by Zikkaron by @amanhij. Different architecture, same north star.

Yadgar stands on a great deal of open-source work — models, databases, frameworks, and tools. Full credits and acknowledgments (including Tom Aarsen, whose sentence-transformers + Ettin reranker underpin our retrieval): docs/reference/tributes.md.

License

Apache 2.0. See LICENSE.

Third-party licenses (key dependencies)

• SurrealDB (default storage backend) — Business Source License 1.1 (BSL). The Additional Use Grant permits embedded use; yadgar bundles + operates SurrealDB as a single-tenant per-deployment store, which falls under that grant. Non-compliant trigger: offering a hosted multi-tenant managed yadgar service exposing the SurrealDB API directly to third-party customers — get a commercial SurrealDB license or migrate to Postgres + pgvector if that becomes a goal.
• surrealdb Python SDK — Apache 2.0 (separately licensed from the server).
• embed / rerank models — sentence-transformers/all-MiniLM-L6-v2, cross-encoder/ettin-reranker-32m-v1 (CE primary), cross-encoder/ettin-reranker-68m-v1 (fallback), Alibaba-NLP/gte-reranker-modernbert-base (rollback), cross-encoder/ms-marco-MiniLM-L-6-v2, cross-encoder/nli-deberta-v3-small: Apache 2.0 weights via Hugging Face.
• Benchmarks (benchmarks/): scripts Apache 2.0, but datasets carry their own licenses — LoCoMo is CC BY-NC 4.0 (non-commercial only); LongMemEval MIT. See benchmarks/README.md.

Full per-dependency audit: docs/reports/audits/license-compliance-audit-2026-05-30.md.