What we tested on 2026-05-28, and why it matters

The short version

We put ai-memory v0.7.0 — the next version of our open-source persistent-memory system for AI agents — through a second round of testing on May 28th, 2026. It passed. The substrate is solid, the documentation now matches the code exactly, and the few issues we found were fixed the same day.

This is the second testing round for v0.7.0. The first round (May 22nd) got 7,321 / 0 / 0 — every test passing. This second round is the “finishing touches” pass: integration testing against a more realistic three-container setup, a one-hour live-usage stress test, a complete documentation audit, and a fresh second-opinion assessment from an independent AI agent.

If you’d like the executive summary, see audience-c-level.md. The deep technical writeup is audience-sme-engineer.md.

What is ai-memory?

ai-memory is software that gives AI assistants a long-term memory. By default, AI assistants forget everything when a conversation ends. ai-memory fixes that — it stores facts, decisions, and context across conversations and across different AI models. It runs on your own computer or server; your memories stay yours.

What we tested

A one-hour live-usage stress test. The maintainer’s actual working AI session, running for a full hour against the new binary. Pass. The daemon stayed alive (in fact, 16+ hours continuously), used about 18 MB of memory throughout (it didn’t leak — that’s the important bit), and answered every request correctly.
A rebuilt three-container “lab” environment. Two simulated AI agents (“alice” and “bob”) and a shared database, exercising a realistic deployment where multiple AI agents share infrastructure. Pass. All three containers came up healthy.
A full Postgres + Apache AGE regression test. 8,028 automated tests against the new three-container setup. Mostly pass — 9 failures, all closed. Of those 9: 5 were build-system race conditions (not real bugs — cleared on re-run); 4 were real test-discipline bugs (the tests assumed a clean database but other tests left state behind). We filed those 4 as issue #1381, built a tiny per-test isolation helper as PR #1382, and all 4 now pass.
A complete sweep of the documentation. Over the past month the product’s numbers changed (e.g., the command-line tool grew from 57 commands to 79; the database schema version moved from v49 to v51). About 27 documentation files still cited the old numbers. We hunted down every stale number with three different audit passes (literal text search, then structural code-graph search, then fault-injection on the “guard rail” tests). About 55 places needed fixing; all are fixed. Closes documentation- drift issues #1197 and #1198.
A scholarly citation. The product’s design has independent conceptual lineage with a published research framework called CoALA (Cognitive Architectures for Language Agents, Princeton, February 2024). We added an explicit citation acknowledging the prior art, with a mapping document showing how our primitives line up. We deliberately did NOT rewrite our roadmap to claim CoALA-compliance — just citation discipline.
A second-opinion assessment from a fresh AI agent. We dispatched an independent AI agent (same model, fresh session, no memory of building any of this) to run its own structural review of v0.7.0 from scratch. That assessment is in-flight as we write this; the report will land under docs/v0.7.0/heterogeneous-ai-nhi-assessment/.

What issues did we find?

Five GitHub issues touched: #1197 + #1198 (documentation drift, both CLOSED), #1378 (the optional auto-installer for the Codex CLI rejects TOML config; substrate’s manual TOML config works fine, so this only affects the optional installer; OPEN, not blocking ship), #1381 + #1382 (the 4 test-discipline failures + the fix PR; CLOSED by the fix).

Why this matters

For someone installing v0.7.0 next week, this campaign is the difference between “the docs sort of describe the product” and “the docs describe the product exactly, every number, every schema version, every command count.” That mattered enough to do three audit passes until zero defects remained.

The 16-hour stress-test result is the other load-bearing signal: a memory substrate that leaked even a little would balloon under a full working day’s use; ours stayed at ~18 MB the whole time.

What happens next

The technical work for v0.7.0 is essentially done. What’s left is:

Operator review and merge of the three open PRs (#1379 docs/Pages drift, #1380 CoALA citation, #1382 test isolation helper).
The 24-hour dogfood loop (a longer version of the 1-hour loop above; operator-driven).
The actual release cut, which is operator-signed-off, not automated. AI agents do the engineering, but the final “this version is now official” call belongs to a human.

In one sentence

We ran the full integration test suite + a 1-hour live-stress test

a complete documentation audit + a scholarly-citation discipline pass on v0.7.0, found five issues, fixed the four blocking ones the same day, and got zero remaining defects across three independent audit lenses — which is what “ready to ship” means around here.