Reference Repository Analysis: Ori-Mnemos and Zikkaron

Retrieval, decay, and learning patterns from two open-source memory systems, cataloged for adoption into Signet.

Reference: references/Ori-Mnemos/ and references/Zikkaron/ in the Signet monorepo.

1. Ori-Mnemos

TypeScript, approximately 12K LOC, 579+ tests.

Architecture

Six-layer stack, bottom to top:

1. Markdown on disk with wiki-link graph structure.
2. Knowledge graph: wiki-links, ACT-R decay, spreading activation, zone-based partitioning.
3. Four-signal fusion: semantic similarity + BM25 keyword + PersonalizedPageRank + warmth, combined via score-weighted reciprocal rank fusion (RRF).
4. Dampening pipeline: gravity, hub, and resolution corrections applied post-fusion.
5. Retrieval intelligence: Q-value reranking, co-occurrence learning, stage meta-learning.
6. MCP server: 16 tools, 5 resources.

Key Patterns

1.1 Q-Value Reranking (memRL-inspired)

Notes earn Q-values from session outcomes via exponential moving average (alpha=0.1).

Phase B reranking blends z-normalized similarity with Q-value, adding a UCB-Tuned exploration bonus for under-explored notes. Cumulative bias cap at MAX=3.0 with compression=0.3. Exposure-aware correction prevents popular notes from dominating indefinitely.

Key files: src/core/qvalue.ts (214 LOC), src/core/rerank.ts (146 LOC).

1.2 Co-Occurrence Edges (Hebbian Learning)

Notes retrieved together in the same session grow edges. Normalization uses NPMI (normalized pointwise mutual information). GloVe power-law frequency scaling prevents common terms from inflating co-occurrence scores. Ebbinghaus decay with strength accumulation governs edge lifetime. Per-node Turrigiano homeostasis prevents hub notes from absorbing all edge weight.

The combined wiki-link and co-occurrence graph feeds Personalized PageRank (HippoRAG-style, alpha=0.5).

File: src/core/cooccurrence.ts (150 LOC).

1.3 Stage Meta-Learning (LinUCB Contextual Bandits)

Each pipeline stage is wrapped in a LinUCB bandit. An 8-dimensional query feature vector drives three-way decisions: run, skip, or abstain. Cost-sensitive thresholds prevent expensive stages from firing on low-value queries. ACQO two-phase curriculum ramps complexity.

File: src/core/stage-learner.ts (150 LOC).

1.4 Intent-Aware Query Routing

Classifies queries into four intents: episodic, procedural, semantic, decision. Classification uses heuristic pattern matching (30+ regex patterns). Each intent type has its own signal weight profile and split weights across title, description, and body fields.

1.5 ACT-R Vitality Model

Seven-factor vitality score:

1. ACT-R base-level activation
2. Metabolic rate (space-dependent)
3. Structural stability boost
4. Access frequency saturation
5. Revival spike
6. Spreading activation boost
7. Bridge protection floor (Tarjan articulation points at >=0.5)

Three memory spaces with distinct decay rates:

1.6 Post-Fusion Dampening

Three stages applied after signal fusion:

1.7 Recursive Memory Harness (RMH)

PPR with exploration-tuned alpha=0.45. Sub-question decomposition with convergence detection. Every retrieval reshapes the graph by updating co-occurrence edges and Q-values.

1.8 Benchmarks

LoCoMo results: 44.7% Recall, 20.8% F1, 32.4% MRR, 63.5% AnsF1.

2. Zikkaron

Python, 26 subsystems, 969 tests.

Architecture

Five cohesive layers:

1. Core Storage and Retrieval: SQLite + vector indexing.
2. Memory Dynamics: thermodynamics, reconsolidation, predictive coding.
3. Consolidation and Organization: astrocyte pool with domain-specialized workers.
4. Knowledge Structure: knowledge graphs, causal discovery, cognitive maps.
5. Frontier Capabilities: Hopfield networks, HDC encoding, rules engines, CRDT sync.

Key Patterns

2.1 Heat-Based Thermodynamic Decay

Every memory carries heat (0.0-1.0), surprise (0.0-1.0, computed as 1 - max_similarity), and importance (heuristic). DECAY_FACTOR = 0.95.

Three compression levels by age:

Anchor and protection mechanisms prevent critical memories from decaying. Forgetting follows rate-distortion optimal curves per Toth et al. (2020).

2.2 Reconsolidation on Retrieval (Nader et al. 2000)

Retrieved memories are evaluated for context mismatch:

mismatch = 0.5 * embedding_distance
         + 0.2 * directory_distance
         + 0.15 * temporal_distance
         + 0.15 * tag_divergence

Three outcomes based on mismatch score:

Tracks plasticity (increases on access, 6-hour half-life) and stability (increases with successful retrieval).

2.3 Predictive Coding Write Gate

Only stores information that violates expectations. WRITE_GATE_THRESHOLD = 0.4 (40% minimum surprisal required). Task continuity discount (WRITE_GATE_CONTINUITY_DISCOUNT = 0.15) lowers the threshold when working on the same task. Bypass keywords exist for errors, decisions, and architecture changes. Adaptive gating tracks the last 10 stored memories.

2.4 Hippocampal Replay for Context Compaction

PreCompact hook drains working state into a checkpoint containing: current_task, key_decisions, files_being_edited, active_errors, custom_context, and epoch. PostCompact hook reconstructs from checkpoint + anchored memories + hot memories + recent actions + predictions. Micro-checkpointing fires every N tool calls (default 50).

2.5 Astrocyte Pool (Background Consolidation)

Runs on idle timeout (default 300s). Domain-specialized workers handle code, decision, error, and dependency processing. The pool runs:

• Thermodynamic decay
• Entity extraction
• Knowledge graph building
• Duplicate merging
• Causal discovery (PC algorithm)
• Sleep replay (random memory pair comparison)
• Louvain clustering

2.6 Hopfield Networks (Ramsauer et al. 2021)

attention = softmax(beta * X * query)

Mathematically equivalent to transformer single-head attention. Fast, no training required. HOPFIELD_BETA = 8.0.

2.7 Typed Knowledge Graph

Bi-temporal model (event_time + record_time). Nine relationship types:

Confidence scoring on inferred edges. PPR with damping factor 0.85.

2.8 Decision Auto-Protection

Regex detection of decision patterns (“chose over”, “decided to use”, “switched from”, “migrated from”, etc.). Matched memories are auto-tagged, set to protected status, given heat=1.0, and assigned a resistance multiplier of 2.0.

2.9 Structured Profile Extraction

Regex patterns extract structured attributes from free-form memory text: interests, traits, travel history, career, goals. Stored as entity_name/attribute_type/attribute_key/attribute_value/confidence. Hindsight inference derives attributes from observed behavioral patterns.

2.10 Benchmarks

3. Cross-Reference Table

Note: For Supermemory ASMR pattern comparisons (structured extraction categories, temporal timeline reconstruction, multi-lens search, ensemble answering, fact supersession tracking), see section 6.

4. Adoption Priority

5. Benchmark Comparison

Signet numbers are from an 8-question sample (run-full-stack-8, 2026-03-22) and have not yet been validated on the full LoCoMo dataset. Metrics are not directly comparable across systems due to differences in k-values, dataset splits, and evaluation methodology. Initial results show perfect recall at k=10 and strong accuracy, with MRR trailing Zikkaron — likely improvable with reranking refinements on the full dataset.

6. Supermemory ASMR (Agentic Search and Memory Retrieval)

Published circa March 2026. Claims ~99% on LongMemEval-s.

Architecture

Multi-agent orchestration pipeline replacing vector search entirely:

1. Ingestion: 3 parallel reader agents (Gemini 2.0 Flash) extract structured knowledge across 6 categories: personal information, preferences, events, temporal data, updates, assistant info.
2. Retrieval: 3 parallel search agents with specialized foci: direct facts, contextual implications, temporal timeline reconstruction.
3. Answering: 8-12 specialized prompt variants running in parallel, with aggregator LLM for consensus.

Results

Critical Assessment

Methodology concern: Run 1’s 98.60% marks a question correct if ANY of 8 independent prompt variants gets the right answer. This is coverage, not single-system accuracy. Run 2 (97.20%) with majority-vote aggregation is the more honest metric.

Cost: 15-19 LLM calls per query (3 search + 8-12 answer + 1 aggregator). Not viable for production use at scale without significant cost reduction.

No embeddings: Uses LLMs as the search engine rather than vector similarity. Eliminates the “semantic similarity trap” for temporal data but trades compute cost for retrieval quality.

Patterns Worth Adopting

Adoption Priority

This document informs amendments to the desire paths epic (DP-9 through DP-15) and new stories DP-16 through DP-20.