Twelfth page in Operation Scarecrow. If you haven't read the argument page and the four memory-type pages yet, start there. Source citations at the foot of this piece.
Everything covered so far in this section is about where memory lives — which of the four tiers a given fact, precedent, or skill belongs in. This page is about a separate axis: how that memory actually gets pulled into an agent's reasoning at the moment of use. Four sub-questions, each with real research behind it.
RAG versus structured retrieval: a routing answer already exists, with numbers behind it
The tradeoff is clean and well established: similarity-based retrieval is cheap, fast, and blind to relationships; structured traversal over a knowledge graph is auditable and handles multi-hop questions but is expensive to build and keep current. Neither wins universally — simple factual lookups favor plain retrieval, relationship-driven and explainability-sensitive questions favor the graph — which is exactly the routing problem the end-state architecture page in this section flagged as the piece nobody had built yet. A named answer already exists: Adaptive RAG, where a query classifier routes each incoming question to the pipeline suited to its complexity rather than running everything through one mechanism. Reported gains are real, not marginal — context-graph-grounded retrieval shows up to 35% accuracy improvement over plain retrieval, and as much as 5x improvement in response accuracy when rich metadata is available. One warning from this research is worth carrying forward word for word: a graph is only as good as the schema behind it, and without ownership, the graph goes stale faster than plain retrieved chunks do — an independent, fourth confirmation that a governed store with no active owner decays, arrived at here from a pure retrieval-engineering angle rather than an organizational-theory one.
Weights versus external store: this is where the framework's whole approach gets a real justification
This is the most consequential finding in this page, and arguably in this section. When agents store reflective experience parametrically — fine-tuned into the model's weights — rather than externally, parametric storage actually outperforms external storage on reasoning tasks, and the performance gap grows specifically on tasks requiring transfer to question types the agent hasn't seen before. Fine-tuning also improves compositional, multi-hop reasoning systematically, in a way retrieval alone can't, because retrieval can hand the model a fact but can't hand it a new way of reasoning over facts in general.
That's a real capability cost to everything this section has been arguing for. But it comes with a decisive counter-finding on the other side: unlearning research shows memory-based editing preserves what should stay known while struggling to actually remove what shouldn't be, whereas retrieval-based approaches outperform fine-tuning specifically at the removal side of the problem. Put plainly — weights are good at adding durable capability and bad at forgetting; external memory is worse at deep reasoning transfer and good at forgetting cleanly. This section's entire governance model runs on the ability to supersede, demote, or retract a fact once it's been challenged and lost — the reconsolidation mechanism, the override-logging rule, the supersede-don't-edit discipline. None of that survives if the fact was baked into weights instead of kept in a queryable, revisable store. This is the concrete, technical reason external memory has been the implicit assumption underneath this entire framework, worth stating explicitly now rather than leaving it implicit: governability requires giving up some of the reasoning-transfer benefit fine-tuning would otherwise provide. That's a real cost, not a free choice.
| Dimension | Baked into weights | Kept in an external store |
|---|---|---|
| Reasoning transfer to unseen question types | Better — the gap grows specifically here | Worse |
| Compositional, multi-hop reasoning | Improves systematically | Retrieval alone can't hand the model a new way to reason, only a fact |
| Cleanly retracting or unlearning a fact | Largely unsolved — editing preserves what should stay known but struggles to remove what shouldn't | Outperforms fine-tuning specifically at removal |
| Compatible with this section's governance model (supersession, override-logging, supersede-don't-edit) | No | Yes — the model depends on it |
Weights are good at adding durable capability and bad at forgetting; external memory is worse at deep reasoning transfer and good at forgetting cleanly. Governability costs some capability — a real trade, not a free choice.
Always-in-context versus on-demand: why stuffing context hurts more than it helps
Letta's three-tier core/recall/archival model is the clearest existing framing here, but the sharper finding is what happens when it's ignored: stuffing everything retrievable into context doesn't help and actively hurts, because of the "lost-in-the-middle" effect — a model reasons worse over a large dump of marginally relevant context than over a small amount of highly relevant context, even though the large dump technically contains more correct information. The practical design rule that follows is to keep the always-on layer deliberately small — lightweight orientation only — and let deeper factual, procedural, and historical content enter active context through deliberate retrieval rather than default inclusion.
The freshness warning here deserves to be flagged as a pattern now, not a one-off: separate memory stores can drift apart without anyone noticing, because vector similarity encodes meaning, not recency — nothing about a similarity score tells you whether the fact it's attached to is still true. It's the same failure mode the poisoning research flagged as a provenance gap and the agent-conflict research flagged as a supersession gap, showing up here in the retrieval layer instead: nothing about time or origin is free information a retrieval system gets automatically. It has to be tagged explicitly, or it silently disappears.
Push versus pull across a fleet: gossip protocols and an energy-decay model for facts
Gossip-protocol research frames this as a spectrum rather than a binary: push propagates updates aggressively from agents that already know something, pull lets agents that don't know request it, and push-pull hybrids reduce worst-case latency while limiting how far different agents' states can drift apart from each other. One system worth naming directly, HyphaeDB, frames agent memory as a living topology where facts carry something like an energy value that decays over time unless reinforced, with promotion between abstraction layers happening through what the paper calls emergent consensus. Durability through reinforcement, decay without it — the same mechanism this framework built from cognitive science and organizational theory, arrived at independently again, this time from distributed-systems and gossip-protocol theory, a field with no obvious connection to either of the other two.
The pattern worth naming across all of this research now
Five separate threads — cognitive-science consolidation, Anthropic's Dreaming, the poisoning research's provenance requirement, the agent-conflict research's temporal supersession, and now gossip-protocol energy decay — have each independently converged on the same two-part idea: durability has to be earned through reinforcement across genuinely separate sources, and origin plus time has to be tracked explicitly or it's lost. That's no longer a claim this section is making once, from one theoretical angle. It's the single most-repeated, most-independently-confirmed finding across the entire Operation Scarecrow research effort to date.
The decision framework nobody's built yet
How an org should actually decide, case by case, whether a given piece of knowledge belongs baked into weights or kept in a revisable external store still doesn't have an answer — the tradeoff is named clearly now, but no decision framework for applying it exists yet in what this research has found. Adaptive RAG's query classifier is also not free — it's an added inference step in front of every query, the same cost this section already flagged for the routing layer in the end-state architecture, and nothing found here prices out how expensive that classification step gets at real fleet volume.
Sources. RAG versus knowledge-graph tradeoffs and Adaptive RAG: "Knowledge Graph vs RAG: When Each One Wins (2026)," Atlan, atlan.com, and "RAG vs Knowledge Graphs: Which Performs Better for Enterprise AI?," techment.com. Fine-tuning versus retrieval, parametric versus external memory, and unlearning: "Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMs," researchgate.net, and "Knowledge Editing for Large Language Models: A Survey," ACM Computing Surveys, dl.acm.org. Always-in-context versus on-demand retrieval, the lost-in-the-middle effect, and freshness as a silent failure mode: "AI Agent Memory vs Retrieval: Why You Need Both," Redis, redis.io, and "Context Management vs Memory Management in AI Agents," Atlan, atlan.com. Push/pull propagation and gossip protocols: "A Gossip-Enhanced Communication Substrate for Agentic AI: Toward Decentralized Coordination in Large-Scale Multi-Agent Systems," arxiv.org/pdf/2512.03285. HyphaeDB: "HyphaeDB: A Living Knowledge Topology for Agent-First Memory," arxiv.org/pdf/2606.28781.