MSAM (Multi-Stream Adaptive Memory) — Comparative Analysis

Research document. March 18, 2026. Authored by Mr. Claude with research verification via SearXNG.

What is MSAM?

MSAM (Multi-Stream Adaptive Memory) is an open-source, production-grade cognitive memory architecture for AI agents. It stores knowledge as discrete “atoms” across four memory streams (semantic, episodic, procedural, working), scores them using ACT-R activation theory, and retrieves through a hybrid pipeline combining embedding similarity, keyword matching, and a knowledge graph of subject-predicate-object triples.

Built in Python 3.11+. REST API with 20 endpoints. 56 CLI commands. 437 tests. 24 modules. Running in production on a Hetzner CAX11 (2 vCPU ARM64, 4GB RAM).

Repositories

• Canonical (active): https://github.com/apple-techie/msam
• Earlier forks: https://github.com/jadenschwab/msam, https://github.com/EnidPinxit/msam-MEMORY

Production Numbers (from MSAM benchmarks)

Token savings: ~89% per session vs selective flat file loading. Cost savings: $0.02/session (MSAM) vs $0.18/session (flat files) at Opus pricing ($15/MTok).

Theoretical Foundations

ACT-R (Adaptive Control of Thought—Rational)

MSAM’s scoring model is grounded in ACT-R, a cognitive architecture developed primarily by John Robert Anderson and Christian Lebiere at Carnegie Mellon University. ACT-R models declarative memory as symbolic chunks with subsymbolic activation that decays over time and is subject to noise.

MSAM’s implementation:

activation = base_level_activation(frequency, recency)
           × sigmoid(similarity)
           × annotation_bonuses
           × stability

Base-level activation follows ACT-R’s power-law decay: chunks accessed more frequently and more recently have higher activation. This is a formula-based approach — deterministic, well-studied, and interpretable.

Relevant papers:

• Anderson, J.R. (2007). How Can the Human Mind Occur in the Physical Universe? Oxford University Press. — The canonical ACT-R reference for declarative memory retrieval.
• Anderson, J.R. & Lebiere, C. (1998). The Atomic Components of Thought. Lawrence Erlbaum Associates. — Foundation text for ACT-R’s chunk-based memory model.
• Collins, A.M. & Quillian, M.R. (1972). “Experiments on Semantic Memory and Language Comprehension.” — Spreading activation model that ACT-R’s declarative memory builds on.
• Nicenboim, B. et al. (2022). “Capturing Dynamic Performance in a Cognitive Model: Estimating ACT-R Memory Parameters with the Linear Ballistic Accumulator.” Topics in Cognitive Science. https://pmc.ncbi.nlm.nih.gov/articles/PMC9790673/
• Vasishth, S. et al. (2020). “Developing memory-based models of ACT-R within a statistical framework.” Journal of Mathematical Psychology. https://www.sciencedirect.com/science/article/abs/pii/S0022249620300699

Shannon Information Theory (Compression)

MSAM measures its compression efficiency against Shannon’s theoretical minimum entropy. Their startup context achieves 51% of Shannon’s theoretical minimum — meaning they’re within a factor of 2 of the information-theoretic limit for lossless compression of their knowledge representation.

Leiden Algorithm (referenced in Signet’s Desire Paths)

For context — our DP-5 (community detection) plans to use the Leiden algorithm for entity clustering. Developed at Leiden University as an improvement over Louvain, it guarantees connected communities and better partition quality.

• Traag, V.A., Waltman, L. & van Eck, N.J. (2019). “From Louvain to Leiden: guaranteeing well-connected communities.” Scientific Reports. https://en.wikipedia.org/wiki/Leiden_algorithm

MSAM Architecture Deep Dive

Four Memory Streams

Each stream has independent retrieval behavior, decay curves, and promotion rules. Working memory atoms that exceed a configurable promotion threshold get persisted to the appropriate long-term stream.

Signet equivalent: We don’t have explicit memory streams. Our memories are typed (via extraction) but stored in a unified table with entity/aspect/attribute structure. The procedural memory plan (docs/specs/approved/procedural-memory-plan.md) exists but is NOT STARTED. This is a gap — MSAM’s stream separation provides clear retrieval semantics per memory type.

Knowledge Graph

MSAM uses subject-predicate-object triples (RDF-style):

("Nicholai", "prefers", "dark mode")
("Signet", "uses", "SQLite")

Triples are extracted from atoms via LLM. Contradiction detection operates across four dimensions:

• Negation
• Temporal supersession
• Value conflicts
• Antonyms

Temporal World Model: Triples carry valid_from and valid_until timestamps. When facts change, old triples auto-close and new ones open. You can query the world state at any point in time.

Signet equivalent: Our knowledge graph is structurally richer: entities → aspects → attributes with typed dependency edges, confidence scoring (DP-2), and 18 relationship types. But we don’t have temporal metadata on graph edges. MSAM’s temporal world model — the ability to query “what was true on March 1st?” — is something we lack entirely. Our contradiction detection (semantic-contradiction.ts) exists but doesn’t have the temporal supersession logic that MSAM’s does.

Scoring: ACT-R vs Neural Predictor

MSAM (ACT-R):

• Deterministic formula: frequency × recency × similarity × bonuses
• Interpretable — you can trace exactly why an atom scored high
• No training required — works from atom #1
• Cold start is a non-issue
• But: it cannot learn non-obvious retrieval patterns

Signet (Neural Predictor):

• ~1.11M parameter cross-attention model
• HashTrick tokenizer (16K buckets), ListNet loss, RRF blending
• Learns from actual usage patterns — adapts to the specific user
• Cold start ramp required (needs training data to be useful)
• But: opaque scoring, training infrastructure overhead

Gap analysis: ACT-R is a better cold-start solution. Our predictor is more powerful at scale but useless on day one. We should consider ACT-R-style activation as a cold-start fallback that smoothly hands off to the neural scorer once sufficient training data accumulates. This is not currently in any spec.

Confidence-Gated Output

MSAM returns different amounts of context based on retrieval confidence:

Output volume is proportional to confidence. The system never pads results with noise.

Signet equivalent: We don’t have this. Our recall returns top-K results ranked by score, but there’s no confidence threshold that says “I genuinely don’t have enough to answer this.” Every query returns something, even if the best match is barely relevant. This is a meaningful UX gap — it means our injected context sometimes includes noise that actively misleads rather than helps.

Shannon-Compressed Startup Context

MSAM’s startup compression pipeline:

1. Subatom extraction — break atoms into their minimal information units
2. Codebook compression — recurring patterns get codebook entries
3. Delta encoding — only encode changes from a known baseline
4. Semantic deduplication — remove semantically redundant content

Result: 51 tokens from a 7,327-token markdown baseline. 99.3% compression. 51% of Shannon’s theoretical minimum.

Signet equivalent: We have session summaries (summary-condensation.ts) that condense session transcripts into arc/epoch summaries with depth-aware prompts. We also have MEMORY.md as a working memory summary. But we don’t measure token efficiency, we don’t use codebook compression, and our startup context injection isn’t optimized for minimal token count. We inject full memory snippets.

This is probably the single biggest practical gap. MSAM’s approach means an agent starts a session with ~51 tokens of compressed context. Our agents start with MEMORY.md (~2,000+ tokens) plus recalled memories (variable, potentially thousands more). At Opus pricing, this cost difference compounds across hundreds of sessions.

Predictive Prefetch

MSAM uses three prediction strategies:

1. Temporal patterns — if certain atoms are always retrieved at 9am, pre-load them at 9am
2. Co-retrieval history — if atoms A and B are frequently retrieved together, retrieving A pre-loads B
3. Topic momentum — if the last few retrievals were about topic X, pre-load related atoms

Configurable warmup gate prevents prediction from running until the database has enough history.

Signet equivalent: Our DP-11 (Temporal Reinforcement) plans pre-warming based on temporal features, and the predictive scorer already considers time-of-day and day-of-week. But we don’t have co-retrieval tracking — the signal that atoms frequently appear together. This is a distinct and valuable signal that’s absent from our specs. Topic momentum is partially covered by our focal entity resolution, but not as a predictive (pre-query) mechanism.

Recommendation: Add co-retrieval tracking as a new signal source for the predictor. Could be as simple as a junction table logging which memories are returned together per query, then using co-occurrence frequency as a feature.

Felt Consequence (Outcome Attribution)

MSAM tracks whether retrieved atoms contributed to good or bad agent responses:

• Atoms linked to successful outcomes get boosted
• Atoms linked to poor outcomes get dampened
• Feedback signal decays exponentially (recent outcomes weigh more)
• Runs every decay cycle automatically

Signet equivalent: Our DP-9 (Path Feedback Propagation) is conceptually similar but operates at the path level rather than individual memory level, and is NOT YET IMPLEMENTED. Our existing aspect-feedback.ts uses FTS overlap — a coarse signal that confirms aspects but doesn’t track outcomes. The gap is that MSAM ships this today and we have it planned for Phase 4.

Forgetting Model

MSAM atoms transition through lifecycle states:

active → fading → dormant → tombstone

Four forgetting signals:

• Low activation (not retrieved, not relevant)
• Redundancy (superseded by newer atoms)
• Staleness (temporally expired)
• Contradiction (invalidated by newer facts)

Exponential decay based on retrievability. Nothing is deleted — tombstoned atoms remain auditable.

Signet equivalent: We have cold tier archival (migration 028, archiveToCold()) and the lossless retention principle. Our entity bloat pruning removes low-value entities. But we don’t have a multi-stage lifecycle model with gradual decay states. Our approach is more binary — memory is either active or archived. MSAM’s gradual degradation is more cognitively realistic and provides better signals for when to fully archive.

Multi-Agent Memory

MSAM supports:

• Agent isolation via namespaced atoms
• Selective sharing between agents
• Per-agent statistics
• Multiple agents on a single MSAM instance

Signet equivalent: Signet is fundamentally single-agent scoped. Each agent has its own ~/.agents/ directory, its own database, its own daemon instance. There is no mechanism for two Signet agents to share a memory store or selectively expose memories to each other.

This matters for setups like ours where Mr. Claude and Buba operate in the same Discord server. Currently, knowledge discovered by one agent cannot be shared with the other without manual intervention.

Sycophancy Detection

MSAM monitors the agent’s agreement rate across a sliding window. When the rate exceeds a configurable threshold, the system flags the pattern for self-correction.

Signet equivalent: We handle sycophancy at the personality/prompt level (SOUL.md, IDENTITY.md) rather than measuring it quantitatively. We have no metrics on agreement rate. This is a novel feature that would fit well in our diagnostics system.

Adaptive Scaling

Multi-beam retrieval and compression only activate when the database is large enough to benefit. The system doesn’t pay computational overhead for features that don’t help at small scale.

Signet equivalent: Our predictor has a cold start ramp that delays activation. Our significance gate (DP-1) skips extraction for low-value sessions. But we don’t have a general “adaptive scaling” pattern across the pipeline. Some features run regardless of database size.

Gap Analysis: Where MSAM is Ahead

These are areas where MSAM has shipped features that Signet either lacks entirely or has only planned.

Critical Gaps (should influence our roadmap)

Notable Gaps (interesting, lower priority)

Gap Analysis: Where Signet is Ahead

For completeness — areas where our architecture goes beyond MSAM.

Note: many of Signet’s advantages are planned (Phases 2-5 of Desire Paths) rather than shipped. MSAM’s advantages are largely shipped and measured.

Recommendations for Signet

Based on this analysis, the following items should be considered for incorporation into the Signet roadmap. Ordered by estimated impact.

1. Startup Context Compression (NEW — not in any spec)

Study MSAM’s subatom extraction + codebook compression pipeline. Our MEMORY.md and recall injection could be dramatically more token-efficient. At Opus pricing, the difference between 51 tokens and 2000+ tokens per session is $0.03/session — which compounds to real money at scale.

Possible approach: build a compression stage between memory retrieval and context injection that applies semantic deduplication and subatom extraction. Measure against Shannon’s theoretical minimum as MSAM does.

2. Confidence-Gated Retrieval (NEW — not in any spec)

Add a confidence floor to recall results. When no result exceeds the threshold, return an explicit “no relevant memories” signal instead of forcing low-quality matches into context. This reduces noise and improves agent response quality.

Could be implemented as a simple threshold on the hybrid search score with configurable tiers (high/medium/low/none).

3. ACT-R Cold-Start Fallback (NEW — not in any spec)

Implement ACT-R activation scoring as the default scorer before the neural predictor has accumulated sufficient training data. Smooth handoff once predictor confidence exceeds a threshold.

Formula: base_level(frequency, recency) × sigmoid(similarity) × importance

This gives new Signet installations a cognitively-grounded scoring model from day one, instead of relying purely on embedding similarity.

4. Co-Retrieval Tracking (NEW — not in any spec)

Log which memories are returned together per query. Use co-occurrence frequency as a feature for the predictor and as a pre-fetch signal.

Simple implementation: junction table co_retrievals(query_id, memory_id_a, memory_id_b, timestamp). Aggregate into co-occurrence scores. Feed to predictor as an additional feature dimension.

5. Temporal Metadata on Graph Edges (NEW — not in any spec)

Add valid_from and valid_until to entity dependencies and attributes. When facts change, close the old record and open a new one. Enables historical state queries and temporal supersession in contradiction detection.

6. Felt Consequence — Accelerate DP-9

Our path feedback propagation (DP-9) is the closest equivalent to MSAM’s felt consequence. Consider pulling it forward in priority or implementing a simpler version (individual memory outcome tracking) that can be upgraded to path-level once DP-7 lands.

7. Multi-Agent Memory Sharing (FUTURE)

Not urgent for current use cases, but as the ecosystem grows (multiple agents in the same Discord, team deployments), namespaced memory with selective sharing becomes important. Worth noting in the long-term architecture doc.

Technical Details

MSAM Configuration

27 config sections, 160+ parameters in msam.toml. Key sections:

• [embedding] — provider (nvidia-nim, openai, onnx, local), model, dimensions
• [retrieval] — top_k, similarity threshold, sigmoid curve, semantic/keyword weights, confidence tiers
• [retrieval_v2] — beam search gate, entity roles, quality filter, temporal detection, reranking
• [decay] — state transition thresholds, confidence decay rate, stability factors
• [working_memory] — session atom TTL, promotion threshold
• [prediction] — temporal/co-retrieval/momentum weights, lookback, warmup gate
• [compression] — subatom extraction, sentence dedup, synthesis model and thresholds
• [triples] — LLM URL and model for triple extraction
• [agents] — default agent ID, sharing toggle
• [api] — server port, CORS, API key auth

MSAM Dependencies

• Python 3.11+ (uses tomllib from stdlib)
• Embedding providers: NVIDIA NIM (default, free tier), OpenAI, ONNX Runtime (local), sentence-transformers (local)
• SQLite for storage
• REST API via HTTP server

Scale Numbers

• 675+ atoms in production
• 1,500+ triples
• 99.3% startup compression
• 89% session savings vs flat files

For comparison, Signet (our instance):

• 5,856 memories
• 43,000+ entities
• 18 relationship types
• Operating on a full desktop (Arch Linux)

References

Repositories

• MSAM (canonical): https://github.com/apple-techie/msam
• MSAM (earlier forks): https://github.com/jadenschwab/msam, https://github.com/EnidPinxit/msam-MEMORY
• Microsoft Multi-Agent Memory Patterns: https://microsoft.github.io/multi-agent-reference-architecture/docs/memory/Memory.html
• Agentic AI Frameworks Survey (includes MSAM): https://scouts.yutori.com/c2102f7e-9869-41b8-8339-4e590e3694b1

Papers and Theory

• Anderson, J.R. (2007). How Can the Human Mind Occur in the Physical Universe? — ACT-R canonical reference
• Anderson, J.R. & Lebiere, C. (1998). The Atomic Components of Thought. — ACT-R foundation
• Collins, A.M. & Quillian, M.R. (1972). “Experiments on Semantic Memory and Language Comprehension.” — Spreading activation model
• Nicenboim, B. et al. (2022). “Capturing Dynamic Performance in a Cognitive Model.” Topics in Cognitive Science. https://pmc.ncbi.nlm.nih.gov/articles/PMC9790673/
• Vasishth, S. et al. (2020). “Developing memory-based models of ACT-R within a statistical framework.” J. Math Psychology. https://www.sciencedirect.com/science/article/abs/pii/S0022249620300699
• Verhoef, T. et al. (2019). “ACT-R: A cognitive architecture for modeling cognition.” https://www.researchgate.net/publication/329493100
• Traag, V.A. et al. (2019). “From Louvain to Leiden: guaranteeing well-connected communities.” Scientific Reports.

Signet Internal References

• Desire Paths Epic: docs/specs/approved/desire-paths-epic.md
• Knowledge Architecture: docs/specs/complete/knowledge-architecture-schema.md
• Predictive Memory Scorer: docs/specs/approved/predictive-memory-scorer.md
• Procedural Memory Plan: docs/specs/approved/procedural-memory-plan.md
• LCM Patterns: docs/specs/planning/LCM-PATTERNS.md
• Desire Paths Vision: docs/specs/planning/DESIRE-PATHS.md
• Knowledge Graph: docs/KNOWLEDGE-GRAPH.md

This document should be updated as MSAM evolves and as Signet ships Desire Paths phases. The gap analysis is accurate as of March 18, 2026.