Appendix C — Shared Retrieval and Ranking for the Auto-Surface and Deliberate Recall, and the Memory-Utility Tradition
C.1 The memory-utility tradition
A 2026 line of work — exemplified by MemRL (Zhang et al., 2026, p. 1) — applies reinforcement learning not
to a model's weights but to its memory retrieval: it learns which stored
memories are worth surfacing by treating retrieval as an action and downstream
usefulness as a reward, ranking memories by a learned Q-value. The
appeal is that it keeps the base model frozen — no fine-tuning, so no catastrophic
forgetting — while still improving over time at the level of what gets retrieved.
Concretely, MemRL composites semantic similarity with an EMA-learned Q-value via a
weighted score, with the Q-error decaying as (1−α)^t (Zhang et al., 2026, pp. 4–5, 13).
ENGRAM sits in this tradition. Every node carries a utility_score — a learned
value updated by the same exponential rule, Q_new = Q_old + α(1 − Q_old) — and
ranking composites that learned value with semantic relevance rather than ordering by
relevance alone. What differs is the reward: MemRL learns from task outcome, while
ENGRAM learns from what the agent did with the node once it had it, which is a
signal the substrate can observe directly and attribute without ambiguity. The rule
has no negative term: a retrieval that proves useless leaves a node's utility where it
was, so utility_score is non-decreasing. Demoting unhelpful nodes would need a
penalty the current rule does not have; adding one is a straightforward change for
a later version.
C.2 ENGRAM's utility_score — a learned value, graded by the act it earned
ENGRAM's utility_score is the Q-value analogue, and what it grades is the
cognitive load of the act a node took part in — not how hard the agent looked to
find it. The learning rate α is keyed to the action, in three bands:
| α | actions | what the band means |
|---|---|---|
| 0.15 | derive, supersede, contradict, resolve, register_exemplar |
the node took part in an epistemic act — it became a premise, drove a revision, or anchored a lesson |
| 0.10 | inspect, subgraph, mention, focus, history, citation |
the node was engaged with — read closely, traversed, named in reasoning, or pinned |
| 0.05 | focus_load, unfocus |
loose touch — the node moved with a set rather than being chosen |
When a node is reached several ways, max-α-wins.
Note what the top band contains: no retrieval verb at all. Every member is
something the agent does with a node once it has it. The table is not measuring
attention spent looking; it is measuring what the looking produced. inspect — the
most deliberate way to read a node — sits in the middle band, below derive, because
reading closely and then building on it are different achievements.
Two independent paths award credit, and the second one matters more than it looks. The first is citation: naming a node as a premise in a new derivation. The second is a stop-hook that scans each of the agent's own turns and credits any node mentioned in it, at α = 0.10. The first path alone was the original design and proved too narrow — it misses prose-mention, inline citation, and reasoning that is plainly informed by a node without formally citing it. A node that shapes an argument without being cited in it has still been useful, and the substrate now sees that.
This is the difference between access frequency (which a vector store can track) and epistemic usefulness (which requires knowing what the retrieval went on to do) — the distinction that puts ENGRAM's ranking in the memory-utility tradition rather than the plain-retrieval one.
C.3 The composite ranking — multiplicative amplifiers over a three-stage pipeline
Retrieval in production — for both the auto-surface and deliberate recall — runs one three-stage _search_nodes pipeline:
- Gather — FTS + semantic-embedding merge produces candidate nodes with a base relevance score.
- Composite re-rank — the stage this appendix is about.
- Diversity — an MMR (maximal-marginal-relevance) pass (MMR_LAMBDA = 0.9) trims near-duplicates so the surfaced set is not nine paraphrases of one claim.
The composite (stage 2) uses multiplicative amplifiers, not an additive blend:
composite = relevance × (1 + UTIL_BETA × utility_score)
× (1 + IMP_BETA × normalized_importance)
with UTIL_BETA = 0.10 and IMP_BETA = 0.005 in current production. The design
rationale (Lei, 2026-05-19) is the load-bearing point: relevance is the base
topical signal; utility and importance are amplifiers in orthogonal dimensions, not
co-equal addends. An additive blend (1−λ)·relevance + λ·utility (the retired
form, UTILITY_LAMBDA = 0.3) requires careful λ-tuning to keep utility from drowning
relevance; multiplicative amplifiers preserve relevance's authority and let each
amplifier scale from 0 to a bounded gain (a max-utility node, Q=1, gets a 1.10×
boost). The importance amplifier is homeostatic — normalized_importance is
anchored to the current minimum-queryable importance, so the formula adapts as the
graph grows rather than using a fixed scale.
C.4 Importance, forgetting, and the testing effect
Two further signals interact with the ranking. Importance gates tier eligibility — it is a pre-filter deciding what is even in the queryable set (tier 2) versus faded (tier 3), per the forgetting mechanism (Ch4 §4.5) — and then enters the composite as the second amplifier. And deliberate recall itself strengthens a node (the recall-count / recall-turn signals), so a node that keeps proving useful is both retained longer and ranked higher: the testing-effect shape, where retrieval is not a read-only event but, as the testing effect shows, enhances later retention (Roediger & Karpicke, 2006, p. 249). The result is retrieval that is adaptive on three coupled axes — topical relevance, learned usefulness, and decaying importance — rather than a static similarity sort.
C.5 Scope
Two qualifications, both consistent with the paper's posture:
- The constants are harness-calibrated, not first-principles. UTIL_BETA, IMP_BETA, and MMR_LAMBDA were tuned by an NDCG sweep on a local recall-measurement harness, which set them to 0.10, 0.005 and 0.9 and gained ≈0.05 NDCG over untuned values. They are measured to rank well, which is a stronger footing than the confidence-model constants (Ch7 §7.2, which are not outcome-calibrated) — a useful internal contrast.
- The reward signal is an engagement proxy, not a true outcome: utility is learned from the engagement actions C.2 lists (citation among them), none of which verifies that the retrieved node was right. A fuller outcome-based reward (did the downstream claim survive contradiction?) is a future direction, continuous with the confidence auto-calibration track (Ch7 §7.2).
→ Ch4 §4.7 makes the claim; this appendix is the detail.