LoomLoom
Loom home Download

Two memories that decay,
reinforce and get deleted

Append-only, per repo, measured. A 1,500-token brief rides every dispatch.

NewLesion harness →
Loom's canon surface: 20,746 nodes of this repository placed in a cortical mesh, memories lit blue

Still frame · 20,746 nodes, 40,033 synapses

  • BM25F
  • k1 8
  • RRF_K 60
  • half-life 21 d
  • brief 1,500
  • entries 167
  • p50 14.9 ms
  • assertions 417

Two memories, not one file. THE RECORD says what happened. THE CANON says what is true. Every mechanism inside them earns its own code.

FIG 0.2 Wireframe of a stack of thin plates

THE RECORD

Episodic. docket.jsonl, append-only, zero LLM calls, 19 event kinds already firing.

FIG 0.3 Wireframe cluster of hollow cubes linked by thin edges

THE CANON

Semantic. 167 typed, anchored, heat-ranked claims. 652 links, one pinned.

FIG 0.4 Wireframe of a fanned cascade of plates

THE BRIEF

Delivery. Five blocks under a hard cap, on the first user message of every dispatch.

Refuse the duplicate.
Keep the correction.

Trigram Dice at 0.82 calls two entries the same entry. A set of discriminating literals runs first, and says they are not.

1.0Write gate →
  1. 1.1Sørensen–Dice
  2. 1.2Discriminators
  3. 1.3Reinforce, never replace
  4. 1.4When in doubt, split

Everything cools
on a 21-day half-life.

Ebbinghaus decay with logarithmic reinforcement. A pin is an exemption from time, never a multiplier on it.

2.0Thermal model →
THE CANON's entry reader, full bleed: one claim's heat curve standing at 0.75 after a single recall
  1. 2.1heatOf
  2. 2.2Reconsolidation
  3. 2.3Cooling sweep
  4. 2.4dormant vs coldNow

Four rankers deleted,
each with its table.

A dense leg inside a rank fusion is a query-independent prior, and it wins. BM25F is the one leg left standing.

3.0Retrieval →
  1. 3.1Weighted RRF
  2. 3.2Dense vs sparse
  3. 3.3Cut, then cap
  4. 3.4BM25F over four bags

1,500 tokens, and
an order to drop them.

Five ceilings that sum past the budget on purpose. The last guard re-reads the assembled text and drops whole blocks.

4.0Injection →
The BRIEF pane, full bleed: 258 of 300 tokens ready, 16 omitted — 14 ranked out beyond the 6-row cap, 2 cut for room
  1. 4.1Spend order
  2. 4.2Drop order
  3. 4.36,000-byte check
  4. 4.4Rides the user message

Two verdicts
nobody wanted.

A paired bootstrap over gold sets, a positive control allowed to fail, and a placebo that must read exactly zero.

5.0The bench →
  1. 5.1M1 at rendered rows
  2. 5.2Per-cluster intervals
  3. 5.3Placebo self-test
  4. 5.4Four verdicts

Changelog

The 70-day expiry

Dormancy stopped being a membership test. 167/167 live, R@6 flat at 0.2744 out to +365 d.

A constant deleted

MIN_INJECT_STRENGTH refused 0 of 503 injected rows across 85 queries, so it went.

Detail budget 200 → 140

Rendered rows 2.38 → 3.04 and spend 240.1 → 261.8 of 300, for zero retrieval work.

Two public suites

LongMemEval-S R@10 97.80% and LoCoMo-10 95.96%, on the same retrieve().

View all →
“The canon does not merely fail to beat grep here. It loses, measurably, and the instrument can see it.”
− label import−0.0250 [−0.0404, −0.0041]
“PIN_PREPEND reads HARMFUL / DELETE on every run. Kept anyway, with the reason written down.”
+ ablations+0.0583 [+0.0147, +0.1236]

This repo’s canon today: 168 rows, 294,373 bytes → 167 entries, 652 links, 1 pinned. The honest ledger →

Measured, not remembered.
Available today.

Download for macOS Read the docs

Loom.dmg · macOS 14+ · Apple silicon canon 294,373 bytes