Blog

4 min readfeature

The Memory Encore round: eleven teardowns, four delight surfaces, one arity bug

Eleven product teardowns shaped four surfaces: freshness dates on every rule, a session-start recall line, exact-context show, a visible health card.

  • #memory
  • #engine

The Recital round asked whether the memory loop works — and drove it like a real user until four bugs fell out. This round asked the question that comes after: works, yes — but can you see it? Does a user sitting in a fresh session know their memory was injected? Can they see a rule going stale? Can they answer "is my memory even loaded?" without trusting us?

To answer it honestly we went back to the field — and this time not to the five memory platforms the Recital round tore down, but to the surfaces users actually touch: the coding agents and the second-wave memory systems. Eleven products, every claim sourced, in docs/memory-encore-round-research.md.

What the field proved

Three findings shaped everything we shipped:

1. The recall moment is the most-iterated memory UX in shipping agents. Claude Code's "Recalled 2 memories" line went through four changelog iterations — the notice, then clickable filenames, then a verbose file list. Windsurf, which retrieves memories invisibly, gets "I haven't seen it make much difference" on Hacker News. Visible memory reads as value; invisible memory reads as nothing.

2. Freshness metadata is now table stakes. Claude Code stamps every memory file with a modified date "helping Claude reason about which memories are fresh vs. stale". Copilot goes further: un-refreshed memories expire at 28 days, with A/B numbers behind the discipline (+3% precision, +4% recall for code review; +7% PR merge rate for the coding agent — their memory system carries citations and verifies them just-in-time against the branch).

3. Cursor killed its Memories feature. Auto-extracted, injected-when- relevant, managed in a settings list — and deprecated in favor of file-based rules. The lesson we took: whatever we automate must stay inspectable, or it reads as creep rather than help. (The forum is now full of third-party memory products for Cursor — the demand did not die with the feature.)

What we shipped

Every learned rule now carries its last-confirmed date. The project memory file renders - run tests with bun test · 2026-11-09 — and past 28 unconfirmed days, - … · 2026-09-13 · stale. The date is honest by construction: it is the platform's updatedAt, so a later dream that re-asserts the rule refreshes it. Evidence that keeps arriving stays fresh; evidence that stopped arriving says so. Seeds and legacy rows without a parseable timestamp stay bare rather than carrying a fake date. The memory guide teaches the semantics to the agent: a stale rule is weaker evidence — verify before asserting. PI_MEM0_STALE_DAYS is the knob (0 turns the marker off; dates stay).

A session-start recall line. One dim line at boot — ◈ memory recalled: 14 rules · v17 · dreamed 12m ago — /memory show — printed once per process when the project has learned rules, silent for fresh projects. Our footer-quiet contract keeps the healthy engine state invisible, which is right for a footer and wrong for recall: this line is the affirmative "your memory was injected" signal, and it works with the engine down because it reads the project's own files.

/memory show. Prints exactly what the model receives this session — every tier's content with its provenance line, the injected guide, per-tier char counts, and the injection-budget tail (pruned/head-truncated/OVERFLOW named, never silent). /learning shows the stack's metadata; /memory show shows the payload. full lifts the 60-line display cap — the model always receives the full stack, only the display trims.

The memory health card, visible at last. The audit — last reconcile, duplicates, supersede conflicts, live warnings — has existed since the SM-2 ticket, but only through the API and the doctor. The dashboard Memory screen now renders it under the injection-budget meter: one quiet fact line, one amber or rose badge per live warning. Visible audits build trust; invisible ones read as creep.

And the bug

The teardown's code-map pass found a real one: engineUndreamedNotes() in the health card returned (undreamedNotes ?? 0).length — the imported function's .length, which is its arity, always 2. The reconcile-stale warning has been reporting a phantom "2 undreamed note(s) pending" whenever the 7-day horizon tripped: an empty queue could never clear it, a real backlog always read as 2. It now reports the real benchmark-lane count, with a regression battery that uses five notes so an arity regression cannot pass by accident.

Verified like a user

The memory-recital benchmark re-ran live on the new code — the same real-sessions-then-recall battery, stamps and all — and the PTY lane drove the new surfaces through the real TUI: the recall line at boot, the exact stack through /memory show, all green. The receipts ride the round log.

The honest gaps we did not ship, on purpose: per-fact citation metadata (Copilot's strongest pattern — it needs an extraction-contract change that deserves its own benchtested round) and graph retrieval over memory links (HippoRAG 2's personalized PageRank; A-MEM's evolving note network). Both are now on the record as future lanes with their evidence attached.

Memory that proves itself, round after round — that's the product.

Spotted something wrong or worth adding? The record is the repo — open an issue.

All posts