Polyphony ARC — Agent Memory Atlas

1. Executive Summary

Polyphony ARC is Mininglamp AI's ARC-AGI-3 harness — 6,542 lines of its own Python beside 79,035 vendored, MIT, a single commit dated 6 July 2026 and no tests directory.

Its published ARC-AGI-3 result is 19.80% at $115 — 2 of 21 games won, 59/157 levels — which is the lower end of that leaderboard and well below the two harnesses this report compares it to. Its memory is a compaction: when the context budget runs out, the model writes a handoff summary and a fresh context starts from it. Two design decisions in hs_compaction.py are worth a report. The summary is bounded and its stakes are stated to the writer — "Under 2500 words. This is the next instance's only memory of your thinking." And the successor is told not to trust it: files on disk are rendered under the heading "On-disk files (authoritative; re-read before trusting memory):"

No marks. Both rules are prompt text, and unlike Tycho none of it is machinery.

2. Mental Model

A compaction is a lossy write, and this harness treats it as one.

The model is told the word budget and told what the words are for, which is a different instruction from summarise the conversation: a writer who knows the summary is the successor's only recollection prioritises differently from one who thinks a transcript is still available.

Then the successor is given the summary and, in the same prompt, a list of the files that outlive it — with the files declared authoritative. Where the two disagree, the prose loses.

3. Architecture

Diagram — the compaction boundary is where memory is created, and the design's two rules — a stated word budget and files-over-prose — are both prompt text at that boundary

Diagram source

flowchart TD
%% caption: the compaction boundary is where memory is created, and the design's two rules — a stated word budget and files-over-prose — are both prompt text at that boundary
    S["session with the model"] --> BUD{"context budget hit?"}
    BUD -->|no| S
    BUD -->|yes| CMP["hs_compaction"]

    CMP -->|"'Under 2500 words.<br/>This is the next instance's<br/>only memory of your thinking.'"| HS["handoff summary"]

    S --> FS[("workspace files<br/>written by the model")]

    HS --> P["successor prompt"]
    FS -->|"rendered listing"| P
    P -.->|"'On-disk files (authoritative;<br/>re-read before trusting memory)'"| PREC["files outrank the summary"]

    P --> S2["successor session"]

    NOTE["design comments:<br/>no per-turn trimming —<br/>it destroyed prefix cache<br/>and working memory"] -.-> CMP

4. Essential Implementation Paths

Compaction is the only memory event. There is no store, no record type and no write path other than the summary the model produces at the boundary.

The bound is stated with its reason. "Under 2500 words. This is the next instance's only memory of your thinking." A cap without a reason gets a truncated transcript; a cap with the stakes attached gets a briefing. It is a prompt-engineering choice rather than a mechanism, but it is the right one.

The file listing carries the precedence rule inline. Rather than a separate instruction to prefer files, the authority claim is the heading of the listing — the successor cannot read the file names without reading that the files outrank its memory.

The comments record what went wrong before. Two survive in the source: one noting that an earlier approach "deleted the agent's WORKING MEMORY", and one that per-turn trimming "destroyed prefix cache + working memory" and was therefore removed. Both are failure modes of aggressive context management, and the current design is legible as a response to them.

5. Memory Data Model

None. The unit is prose under a word cap. There is no field for confidence, for provenance, for time, for scope or for a rejected conclusion, and nothing parses the summary after it is written.

This places the system at the same point as Retrodict — the memory contract is a prompt — with one difference in the contract's content. Retrodict asks the model to mark each point checked against the log or still assumed; Polyphony asks for a bounded summary and states which artifact wins a disagreement. Neither is read by code. Retrodict's version says more.

6. Retrieval Mechanics

There is no retrieval. The summary is injected into the successor's prompt in full, and the file listing beside it is a directory rendering rather than a query. Nothing is ranked and nothing is selected, which for a single bounded summary is coherent.

7. Write Mechanics

One write, at the compaction boundary, by the model. No validation, no schema, no size enforcement beyond the instruction, and no diff against the summary it replaces.

8. Agent Integration

The harness wraps a long-horizon ARC-AGI-3 session with a deliberate choice not to trim per turn — the comment above says why — so the working context stays intact until the budget forces a full compaction. The vendored tree is twelve times the size of the harness's own code and was not read for this review.

9. Reliability, Safety, and Trust

Single agent, single workspace, no other principal: scope, audit and review have no consumer here, and their absence is not a defect. What is a defect is that neither of the two good rules is checkable. A summary that runs to 4,000 words, or one that contradicts a file it was supposed to defer to, produces no error and leaves no trace.

The absence of a tests directory matters more than the marks it costs. This is a harness whose entire memory behaviour is emergent from prompt text, published with a benchmark claim and without a single committed assertion about the compaction path.

10. Tests, Evals, and Benchmarks

There is no tests directory. Nothing was run for this review.

The published result is the official scorecard, listed on the ARC Prize community leaderboard as 19.80% on ARC-AGI-3 Public Demo for $115, dated 7 July 2026 and tagged v11.2-swarm. Underneath: 2 of 21 environments won, 59/157 levels, 6,838 actions. Two games finish at 100.00 (lp85-305b61c3 and tu93-0768757b) and eleven score below 5. The 21 per-environment scores average to 19.8029, so the figure recomputes from its own table.

Two qualifications belong beside it. The metric is RHAE — relative human action efficiency — so 19.80% is not a completion rate, and the 59/157 levels is the closer thing to one. And this run covers 21 environments where Tycho and Retrodict cover 25: ft09, r11l, sb26 and sc25 are absent here, so the three numbers are not means over the same set, which matters for anyone below a ceiling. The leaderboard hosts but does not check these — only ARC-AGI-1 and ARC-AGI-2 semi-private results are run and verified by ARC Prize, and "everything else is scored on a public set and self-reported."

A harness at 19.80% with no committed test of its compaction path is not a system whose memory design has been demonstrated. The two prompt lines in section 4 are worth reading because they are well-stated, not because a result vouches for them.

11. For Your Own Build

Tell the summariser what the summary is for. "This is the next instance's only memory of your thinking" changes what a model chooses to keep, at no implementation cost.

Put the precedence rule in the artifact, not beside it. Making files are authoritative the heading of the file listing means the reader cannot skip it.

Leave the failure in the comment. "No per-turn trimming (that destroyed prefix cache + working memory)" is why the next person will not reintroduce it.

12. Open Questions

What does a real handoff summary look like? None is committed, so whether the 2,500-word bound is honoured, and what a compaction drops, is unmeasured.

Is anything carried between games? The summary is within-run. Nothing found here moves a conclusion across runs.

What is in the vendored tree? 79,035 lines, unread for this review; a memory mechanism inside it would not have been seen.

Appendix: File Index

Path	What it holds
`arc_hs/hs_compaction.py`	The compaction, the word bound, the authoritative-files heading, the design comments
`arc_hs/`	The harness's own code, 6,542 lines
vendored tree	79,035 lines, not read

History

2026-08-28 — 9bb384c2… — same commit, second reading, covering the published scorecard: 19.80% on ARC-AGI-3 Public Demo at $115, 2 of 21 environments won, 59/157 levels, 6,838 actions, recomputing from its own per-environment table. The run covers 21 environments where the two harnesses this report compares against cover 25, so the three means are not over the same set. Sections 1 and 10 now carry the result, which the first reading described only as "a benchmark claim". Marks unchanged at none.

2026-08-27 — 9bb384c2… — first reading, MIT, 6,542 lines of harness code beside 79,035 vendored, a single commit dated 6 July 2026, no tests directory. Screened before reading: no auto-run surface, no execution surface, one unpinned dependency surface. Nothing was installed and nothing was run. No marks. The whole memory mechanism is a compaction that asks the model for a handoff summary under 2,500 words and hands the successor that summary beside a file listing headed as authoritative. trust_state is withheld: the files-outrank-memory rule is prompt text with nothing reading it, and no memory carries a status field. tombstone, bitemporal, scope_enforced, audit_log and human_review are absent, and negative_eval cannot be earned — there are no tests. The reading covers the harness's own code; the vendored tree was not read.