&memory.episodic.store/replay`act(action: "store_trace")` — persists a full OS-011 InteractionTrace (trace header + ordered edges) in episodic memory. Validates the payload against the canonical OS-011 §4 schema before writing.
`retrieve(action: "replay")` — fetches a trace (by trace_id or by initial state_hash for procedural cluster lookup) and returns a replay manifest ready for a &body.* provider to execute. Destructive traces are flagged with re_authorization_required: true per OS-011 §4.4.
New types: Graphonomous.Types.InteractionTrace and nested Edge — validating parsers for both JSON strings and decoded maps.
New tables: interaction_traces + trace_edges (migration 2026_04_21_interaction_traces_os011), indexed on body_subtype, initial_state_hash, started_at, and trace_edges.state_before for procedural clustering lookups.
19 new tests covering schema validation, Store CRUD, machine routing, destructive-action flagging, and state-hash lookups. Total: 573 tests, 0 failures (was 554).
Unblocks OS-011 conformance tests 7 (trace_edge_complete), 9–10 (replay state-divergence), and 11 (re-authorization) for downstream &body.* providers.
Enables website/OS FSM emergence, skill crystallization (Graphonomous → FleetPrompt SkillCandidate), and cross-machine replay — the core of the dark-factory closed loop.
Step 2 of the 7-step dark-factory loop (Records InteractionTrace) moves from ❌ to ✅ in STACK_COMPLETION.md.
Embedder upgrade: Switched from all-MiniLM-L6-v2 (384D) to nomic-embed-text-v2-moe (768D, 500M params) — improved semantic recall across all retrieval paths
MCP tool: `trace_evidence_path` — weighted Dijkstra / Yen's K-shortest evidence paths with time-decay edge weighting and bidirectional search
Graph algorithms library (6 algorithms, 72 new tests):
Dijkstra weighted shortest path (Graphonomous.Graph.Algorithms.Dijkstra)
DAG detection + topological sort (Graphonomous.Graph.Algorithms.DAG)
Bipartite matching: Hopcroft-Karp maximum + Hungarian optimal assignment (Graphonomous.Graph.Algorithms.Matching)
Louvain community detection with modularity scoring (Graphonomous.Graph.Algorithms.Louvain)
Incremental SCC maintenance (Graphonomous.Graph.Algorithms.IncrementalSCC)
Triangle counting with clustering coefficient (Graphonomous.Graph.Algorithms.TriangleCounting)
GraphMemBench v2: κ-sensitive synthetic benchmark with 8 tiers (T1–T6 κ-topology, T7–T8 graph algorithms), difficulty knobs, and mixed-κ scenarios
LongMemEval full evaluation (ICLR 2025, 500 questions, oracle split):
92.6% QA proxy accuracy, 98.7% session hit rate, 1.4s mean latency
Topology ablation: ON = 92.6%/98.7% SHR vs OFF = 92.3%/97.9% SHR (+0.3pp QA, +0.8pp SHR)
Abstention accuracy: 96.7% (29/30 correct) via learned ANN-statistics threshold
LMStudio judge backend for automated LongMemEval scoring
PPR experiment: Personalized PageRank tested at w=0.18 and w=0.10; net negative on LongMemEval, flag-gated off (enable_ppr)
455 tests (up from 309), 0 failures
MCP tool count: 28 → 29 (+ trace_evidence_path)
Embedder dimensionality: 384D → 768D (retrieval, consolidation, and all similarity operations)
Retriever: added skip_topology option for ablation testing; tuned SCC limit
LongMemEval retrieval fixes: session-aggregate ranking boost, relative-date parser, preference query handling
Graph algorithms attributed to v0.3.2 in README were actually shipped in v0.3.3
Preference-type queries returning wrong sessions (query-type routing fix)
Relative date references ("last Monday") not resolving correctly in temporal retrieval
Session-aggregate ranking not boosting multi-turn conversation context
MCP tool: `belief_revise` — AGM-rational belief revision with three operations: expand (add + contradiction detection), revise (supersede with provenance tracking), contract (retract + propagate)
MCP tool: `belief_contradictions` — detect contradicting nodes via embedding similarity (threshold 0.75) with content_preview and confidence
MCP tool: `forget_node` — intentional forgetting in three modes: soft (hidden from retrieval, reversible), hard (UtU constant-time delete), cascade (hard + orphan propagation)
MCP tool: `forget_by_policy` — budget-aware hybrid priority pruning; scores nodes by confidence × recency × access × connectivity; supports dry_run with candidate listing
MCP tool: `gdpr_erase` — GDPR Article 17 compliant permanent deletion with audit record (node_type, reason, erased_at); no recovery
MCP tool: `epistemic_frontier` — Wilson score confidence intervals at 95%; returns nodes sorted by information gain with interval_lower, interval_upper, width, evidence_count, access_count
Consolidation Stage 4.5: contradiction detection during consolidation pipeline
Q-value tracking: learn_from_outcome now updates q_values alongside confidence; retrieve_context uses q_values for outcome-weighted ranking
Multi-agent scoping: agent_id metadata field for per-agent node attribution; shared graph with cross-agent discovery via graph expansion
Precondition matching: typed-retrieval now boosts nodes matching query preconditions
17 interactive A/B demos covering all 28 MCP tools (up from 11 demos)
309 tests (up from 240), 120 GraphMemBench scenarios across 5 categories
MCP tool count: 22 → 28 (+ 5 resources unchanged)
Consolidation pipeline: 7-stage → 8-stage (added contradiction detection)
learn_from_outcome response now includes new_q_value, old_q_value, q_update_count in updates array
Demo suite expanded: 6 new demos (belief-revision, intentional-forgetting, epistemic-frontier, causal-chains, qvalue-retrieval, multi-agent-memory), 2 existing demos updated
store_node MCP timeout bug: GenServer.call defaulting to 5s while embedder takes up to 15s; added 30s timeout + exit catch in facade
All demo payloads now match real MCP tool output structures (previously showed fabricated response fields)
3 new node types: temporal (time-indexed observations), outcome (empirical results), goal (durable intent)
11 new edge types: causes, resolves, related_to, part_of, follows, supersedes, depends_on, similar_to, temporal_before, temporal_after, co_occurs
4 new node fields: causal_parent_ids, creation_source, timescale, decay_rate
2 new edge fields: co_activation_count, decay_rate
MCP tool: `delete_node` — remove a node and its connected edges
MCP tool: `manage_edge` — edge lifecycle management (list_all, list_for_node, update, delete)
Store APIs: list_all_edges/0, update_edge/2, delete_edge/1
DDL migrations: 2026_04_02_node_spec_fields, 2026_04_02_edge_spec_fields (ALTER TABLE for existing databases)
Consolidation stages 3-7: prune_weak_edges, strengthen_coactivated, merge_similar_nodes, promote_timescale, generate_abstractions
Timescale system: 4-tier decay (fast/medium/slow/glacial) with promotion based on access frequency
240 tests (up from ~148), including direct MCP execute/2 tests for all 22 tools
Edge weight default: 0.5 → 0.3 (spec-aligned)
Node creation_source defaults to :inference (was nil)
Node timescale defaults to :medium (was nil)
MCP tool count: 20 → 22 (+ 5 resources unchanged)
Backward-compatible: legacy edge types (causal, related) still accepted
access_recency field reference in retrieve_episodic and recent_nodes resource (was silently returning nil, now correctly uses last_accessed_at)
store_edge MCP tool default weight hardcoded to 0.5 in 3 places (now 0.3)
Graph.normalize_edge_attrs default weight was 0.5 (now 0.3)
Graphonomous.update_node/2 argument order bug — attrs were piped as first arg to Graph.update_node/2 instead of node_id (caused LearnFromFeedback correction to crash)
Previous release (20 MCP tools, 3 node types, 5 edge types)