Agent Memory Atlas

A governed memo log inside a coding-agent control plane

aios (rexleimo)

An append-only memo log inside a local coding-agent control plane, where automatic writes wait as candidates for a promotion hard-wired to refuse.

Source rexleimo/aios Analyzed revision 6e2910a9…
Analyzed on 2026-09-30
Archived fork rexleimo--aios
LicenceMIT
Size122,233 lines of JavaScript and TypeScript outside tests; the memo layer, dream lane and ContextDB are about 14,600 of them
Activity1,036 commits on main from 5 contributor identities, 1 March 2026 – 28 September 2026
Tests2,486 test cases in 292 files; 228 in the 32 memo and dream files

Carries 5 of 7 rubric mechanisms. Most systems here carry none or one (44%), and a dash means the mechanism was not found at this commit — not that the system needed it. Each mark is one LLM reviewer's reading of the code at this commit rather than a run of it — known limits.

  • Tombstone
  • Trust state
  • Bi-temporal
  • Scope enforced
  • Mutation audit
  • Human review
  • Negative evals

1. Executive Summary

aios (rexleimo) is a local control plane that wraps coding-agent CLIs — Codex, Claude Code, Gemini, OpenCode and others — with hooks, a planning runtime and a project memory. It is not agiresearch/AIOS. This report reads only the memory: the memo event log, its governance and dream lanes, and the ContextDB session store that unified search reads beside it.

What is notable is the belief model. A memo is an append-only event with a provenance block, a claimStatus, a validAt and optional supersedes links. Correction is an appended event, checked against an ACL on which scope and agent may retire what. An automatic write from a model's verified=yes declaration lands as candidate, which every memo read drops.

What is weak is that the governance is sealed at both ends. The promote, reject and expire verbs are hard-wired to deny, so no candidate is ever promoted. Every agent-facing explicit write — MCP memory_write, the Pi tool, aios memo add — carries no identity and lands verified as local-user.

The agent-facing memo skill contradicts the code on exactly this. It tells the model that memo add "proposes — the entry lands as candidate" (skill-sources/memo/SKILL.md:92-94). The project's own test asserts the opposite: a CLI memo add is verified with trust: local_manual (scripts/tests/memo-provenance.test.mjs:196-225).

Five marks: trust_state, bitemporal, scope_enforced, audit_log and negative_eval. Section 9 names the two withheld, tombstone and human_review, and why.

2. Mental Model

A memory is a memo event: text, refs, a space, a scope, an agent, a claim status, a provenance block and a validity start. It becomes a belief in one of two ways. A write with no runtime identity is verified with provenance local_manual, principal local-user (scripts/lib/memo/storage/provenance.mjs:47-68). A write with a runtime identity is verified only if the identity is human or carries a publish capability; otherwise a shared-scope write is candidate (:71-75).

A caller cannot declare its own status. createMemoEvent honours a supplied claimStatus only when it rides on trusted provenance, which only migration passes. Anything else takes the authority verdict, so a forged verified degrades (scripts/lib/memo/storage/events-write.mjs:62-75). Environment variables do not count as identity; a test poisons AIOS_RUNTIME_* and asserts the verdict is unchanged (scripts/tests/memo-authority-env.test.mjs:40-59).

Who passes an identity decides everything. The automatic write after a one-shot turn builds a bare runtime identity with shared: false (scripts/lib/memo/autopilot.mjs:228-235), and aios import does the same (scripts/lib/memo/import-external.mjs:96-106). Those land as candidates. The MCP memory_write handler, the Pi tool and the CLI pass none, so their writes land verified (scripts/memory-mcp-server.mjs:96-115, packages/aios-pi/lib/aios-cli.mjs:87, scripts/lib/cli/dispatch.mjs:398-400).

A belief stops being live by supersession or by age. A later event whose supersedes names it sets invalidAt to its own validAt at read time (scripts/lib/memo/storage/temporal.mjs:112-139). Shared events may retire shared events; private events may retire only the same agent's private events; a shared event retires a private one only when it is a verified promotion (:69-88). Unknown targets are denied at write time and recorded in supersedeDenied (:90-108). Private rows older than 30 days are hidden from recall without deletion (scripts/lib/memo/storage/query.mjs:35-60).

A candidate never becomes a belief. The governance module lists candidates, scans a promoted text for injection, and would append a verified event carrying promotionOf. But authorize returns allowed: false with trusted_authority_unavailable for every action, and hasBrokerReviewAuthority returns false (scripts/lib/memo/storage/candidates.mjs:181-183, :230-237). Each attempt writes a DENY receipt. The comment calls it a broker-reserved seam (scripts/lib/memo/cli/run.mjs:46); no broker exists in the tree.

Diagram — how a memo becomes a belief in aios, the two write authorities, and the promotion that always refuses
Diagram source
%% caption: how a memo becomes a belief in aios, the two write authorities, and the promotion that always refuses
flowchart TD
    A["aios memo add / MCP memory_write / Pi aios_memory_write"] -->|"no runtime identity"| V["verified, trust local_manual, principal local-user"]
    T["one-shot turn declares verified=yes"] -->|"bare runtime identity, shared scope"| C["candidate"]
    T2["one-shot turn failed or not shared"] --> P["agent_private, verified"]
    I["aios import"] -->|"tool identity, no capability"| C
    C -->|"memo reads filter claimStatus"| H["hidden from memo list, search, recall"]
    C -->|"memo candidate promote / reject / expire"| D{"authorize()"}
    D -->|"always allowed: false"| R["DENY receipt appended, candidate stays pending"]
    V -->|"later event with supersedes, same space and scope"| S["invalidAt = successor validAt"]
    S -->|"default read"| X["hidden; --include-invalid or --as-of shows it"]
    P -->|"older than 30 days"| X2["hidden from recall, still on disk"]
    T -.->|"response text also appended"| CDB["ContextDB session events, no status"]
    CDB -.->|"unified search contextdb arm"| U["recallable in later sessions"]

3. Architecture

The memory is Node.js modules invoked by the aios CLI, by client hooks and by two stdio MCP servers. There is no daemon. Everything lives under the workspace's .aios/ state root (scripts/lib/aios/state-root.mjs:42-56).

  • Memo store (scripts/lib/memo/storage/, about 3,000 lines): events in .aios/memo/file/events.jsonl, or one NNNNNNNNNNNN.json per sequence under split/events/. Pinned Markdown per space under file/pinned/. A lockfile serialises writers (lock.mjs).
  • ContextDB (mcp-server/src/contextdb/, TypeScript, about 5,800 lines): per-session l2-events.jsonl and checkpoints, index JSONL files, and a SQLite FTS5 sidecar marked rebuildable (sqlite/schema.ts:75-140).
  • Governance and telemetry: candidate receipts, dream receipts, recall feedback and automatic-write receipts, each an append-only JSONL.
  • Unified search (scripts/lib/search/unified-search.mjs) merges memo, pinned, ContextDB, plans, docs and code for the CLI, the MCP recall tool and per-turn recall.
  • rex-harness is a git submodule at 869b87fe… holding the planning runtime; its tree has no path naming memory or recall, and it was not read.

Deployment and ergonomics

Node 24 and a local filesystem are all it needs; nothing calls a model or a network service to store or recall. No API key is required. The install path is a curl | bash release installer followed by aios init --all, which writes hooks and MCP entries into each detected client. The memo store is JSONL a person can read and repair by hand, and the SQLite sidecar is rebuilt from the JSONL on demand.

4. Essential Implementation Paths

  • Explicit write. aios memo add → handleMemoAddCommand (scripts/lib/memo/cli/commands/events.mjs:53-122) → safety and length checks → appendMemoEvent (scripts/lib/memo/storage/events-write.mjs:159-216), which partitions supersedes against the space's events under the lock. A verified add is also mirrored into a legacy ContextDB workspace-memory session (events.mjs:100-102, scripts/lib/memo/cli/legacy.mjs:87-111).
  • Automatic write. The shell bridge routes a print-mode client call to the one-shot flow (scripts/lib/contextdb/shell-bridge/main.mjs:52-55). After the response, parseMemoryDeclaration reads a trailing <!--memory: verified=yes|no, ...--> block, and only verified=yes calls recordAutomaticMemory (scripts/lib/ctx-agent-core/run.mjs:409-438). That function redacts secrets, runs the content-safety scan, dedups by source ref and appends (scripts/lib/memo/autopilot.mjs:163-282).
  • Session close. autoMemoSessionClose writes a session-close-memory-candidate.json sidecar beside the ContextDB session, never into the memo log (scripts/lib/lifecycle/session-hooks/close.mjs:69-123).
  • Retrieval. searchMemoEvents reads the whole space, drops dream-archived ids, applies temporal, candidate, scope and TTL filters, token-matches, scores and slices (scripts/lib/memo/storage/query.mjs:304-358).
  • Context assembly. The UserPromptSubmit hook calls collectTurnRecall (scripts/lib/planning/user-prompt-submit.mjs:34-45), which queries memory, ContextDB and plans with six hits and 2,000 characters (scripts/lib/planning/turn-recall.mjs:17-20, :218-228), then appends the declaration instruction.
  • Correction. memo add --supersedes, or memo supersede --apply, which re-asserts the winning text as a new event retiring every revision at similarity 0.82 (scripts/lib/memo/cli/commands/supersede.mjs:74-86, temporal.mjs:15, :160-195).
  • Governance. memo candidate list|inspect|promote|reject|expire → decideCandidate (scripts/lib/memo/storage/candidates.mjs:283-406); dream approve, archive, restore and GC → the same shape in scripts/lib/lifecycle/dream/governance.mjs:187-214.
  • MCP. aios-memory exposes memory_recall, memory_write and memory_checkpoint (scripts/memory-mcp-server.mjs:26-62), registered for Codex (scripts/lib/native/emitters/codex-config.mjs:91-100) and Pi (scripts/lib/components/pi/mcp-adapter.mjs:66-71).

5. Memory Data Model

A memo event carries schemaVersion, eventId (memo:<space>:<timestamp>-<uuid8>), storage, space, spaceKey, seq, ts, role, kind, text, refs, scope, agent, claimStatus, provenance, validAt, and optionally supersedes, supersedeDenied, promotionOf, entities, confidence, evidenceRef and turn (scripts/lib/memo/storage/events-write.mjs:80-107). invalidAt and supersededBy are never stored; they are derived per read.

Provenance is a fixed record: trust in runtime_attested, local_manual or legacy_unknown; producerType; principalId; agentId; session, run and activation ids; policyRevision; sourceRef; a sha256 sourceHash; and capabilities (scripts/lib/memo/storage/provenance.mjs:97-134). Rows written before provenance read back as legacy_unknown.

Scopes are project_shared, agent_private and agent_ephemeral, with project, shared and global folded into the first (scripts/lib/memo/storage/normalizers.mjs:46-52). The space is a second key, filtered in collectEvents; in split mode it is also a directory.

ContextDB is a different kind of record. Its events are session transcript lines — prompt, response, error — keyed by sessionId#seq with project and agent from the session, and checkpoints carry a status, summary, next actions and telemetry (mcp-server/src/contextdb/sqlite/schema.ts:78-131). Nothing in it has a claim status or a supersede link.

6. Retrieval Mechanics

Memo retrieval is lexical and computed per query. A row matches when the lowercased query is a substring, or when enough query tokens hit: all of two, half of up to six, a third beyond (scripts/lib/memo/storage/query.mjs:94-123). Tokens are Latin or identifier runs kept whole, ICU word segments, and CJK character bigrams as a fallback (:74-92).

Ranking adds 2 for a text substring hit, 1 for a refs hit, four times a BM25 score normalised by its best case, and an entity score (:188-199). BM25's IDF is computed over the matched set rather than the whole store (:161-186). The entity layer boosts rows whose declared entities hit the query and spreads half of that one hop to rows sharing an entity (:216-274). Recall feedback adds up to 4.5 for rows marked useful and multiplies by 0.6 after five impressions with no use (scripts/lib/memo/storage/feedback.mjs:83-88). Ties inside 0.2 fall to recency.

The optional embedder is a 256-bucket signed hash of the same tokens, not a model (scripts/lib/memo/storage/embedding.mjs:16-45). It can only add rows to the matched set, and a comment in unified search states there is no vector backend.

The trust filter covers one arm of a merged result. Unified search puts ContextDB transcript events beside memo rows (scripts/lib/search/unified-search.mjs:263-321). The one-shot flow appends the full response, up to 8,000 characters, as a ContextDB event before the automatic memo write (scripts/lib/ctx-agent-core/run.mjs:80-88). So the text a candidate holds is recallable in a later session through the ContextDB arm while the memo arm hides it.

Per-turn injection is bounded: six hits, 320 characters per memory, 2,000 in total, placed in the hook's additionalContext after the cached system prefix.

7. Write Mechanics

Writes are deterministic; no model extracts or consolidates anything. The model contributes only its declaration block: whether the turn is verified, which recalled ids it used, and a one-line conclusion (scripts/lib/memo/declaration.mjs:1-57). The automatic memo text is assembled from changed files, the declared conclusion, the task, a clipped result and the outcome (scripts/lib/memo/autopilot.mjs:103-132).

A failed, errored or blocked turn is filed agent_private even when the model declared success (autopilot.mjs:220-223). A duplicate is detected by source ref, not by text. aios import skips facts whose normalised text already exists. Content matching the workspace-memory safety scan is refused on the explicit and automatic paths and at promotion.

Supersession is the only correction. memo add prints a hint when a live fact scores 0.7 or more against the new text and writes nothing (scripts/lib/memo/storage/temporal.mjs:197-220); the writer then passes --supersedes. The memo CLI's verbs are use, gui, storage, space, persona, pin, checkpoint, useful, add, recall, list, search, supersede, candidate, hygiene and report; none deletes an event (scripts/lib/memo/cli/run.mjs:57-189).

The pinned block is the exception to append-only. writePinnedMemo and appendPinnedMemo rewrite file/pinned/<space>.md under the lock, with an optional hash guard against a stale read (scripts/lib/memo/storage/pinned.mjs:89-126). It has no provenance and no status, and MCP memory_checkpoint appends to it.

Operational cost

The explicit write is a synchronous append under a file lock; the automatic write runs after the response is printed, so the turn does not wait on it and a memo is retrievable on the next read. Every search reads the space's whole event file (an in-process parse cache exists) and scores the matched set in JavaScript, so cost grows with the store. The optional embedder hashes every row's tokens per query. Nothing rewrites the store in the background: autodream is opt-in through AIOS_AUTODREAM_AUTO=1, runs preview only, and writes a proposal file (scripts/lib/memo/autodream-auto.mjs:1-30).

8. Agent Integration

Three surfaces reach the model. The UserPromptSubmit hooks for Claude Code, Codex and Grok inject recall and the declaration instruction each turn. The aios-memory MCP server gives Codex and Pi recall, write and checkpoint tools. The Pi extension adds aios_memory_useful and maps each tool to an aios memo command (packages/aios-pi/lib/tools.mjs:25-70).

The model's agency is uneven by surface. Through MCP it writes verified shared memory; through the CLI it can also supersede any shared fact in the space. Through the one-shot wrapper its declaration yields a candidate it cannot promote. The declaration block is requested on every hooked turn, but only the one-shot flow parses it for a write (run.mjs:412); in an interactive native session it is read only for useful-feedback ids (scripts/lib/planning/turn-recall.mjs:178).

Adapting the store to another agent is cheap: appendMemoEvent and searchMemoEvents are plain async functions over files.

9. Reliability, Safety, and Trust

The candidate queue is sealed. Promotion, rejection and expiry all return DENY with trusted_authority_unavailable, and candidate text cannot be listed or inspected (scripts/lib/memo/storage/candidates.mjs:181-237). The suite asserts the denial against a spoofed human identity, eight concurrent promotions and the CLI (scripts/tests/memo-candidate-governance.test.mjs:143-183, :221-236). The result is a write-only queue: every automatic memory from a verified shared turn is withheld forever.

The explicit write surfaces bypass it. handleMemoryWrite calls appendMemoEvent with scope project_shared and no runtimeIdentity (scripts/memory-mcp-server.mjs:96-115). The authority falls to the manual branch: verified, local_manual, principal local-user, capability memo:publish-shared (scripts/lib/memo/storage/provenance.mjs:49-68). The model-written memo is recorded as a person's. The project's own backlog states the rule the MCP server breaks: any non-manual writer must carry a runtime identity (docs/plans/2026-09-08-memo-optimization-backlog.md:126).

The skill misdescribes the gate to the model. skill-sources/memo/SKILL.md:90-98 says memo add lands as candidate and a self-awarded verified is demoted. The code does the reverse for memo add, and scripts/tests/memo-provenance.test.mjs:196-225 asserts it.

Scope is a label, not an identity. --agent and AIOS_AGENT_ID name the reader; any process naming an agent reads its private rows. Supersession is better guarded: the ACL denies cross-scope and cross-agent retirement and fails closed on unknown ids (scripts/lib/memo/storage/temporal.mjs:69-108).

Withheld marks. tombstone: nothing records a rejected value. Supersession is keyed on event ids, the dream lane's tombstone actions are keyed on event ids and never applied (scripts/lib/lifecycle/dream/index.mjs:115-131), and the reject verb is denied. human_review: a queue exists, but no actor can resolve it; a queue nothing drains fails the mark.

Data loss and concurrency. Writers take a lockfile; tests cover concurrent appends (scripts/tests/memo-storage-locking.test.mjs). Nothing deletes a memo event. The ContextDB mirror of a verified add is a second copy that supersession does not reach.

10. Tests, Evals, and Benchmarks

The memo and dream suites hold 228 test cases in 32 files; ContextDB adds 46 in three TypeScript files. CI runs the scripts regression suite (.github/workflows/ci-main.yml:77-98). I read the tests and ran none.

The negative cases are the strongest part. The scope case asserts exact equality to the shared row for another agent (scripts/tests/memo-scope.test.mjs:38-54). The provenance case asserts the active fact is present and the candidate absent (scripts/tests/memo-provenance.test.mjs:106-126). The temporal CLI case matches the new value before asserting the superseded line is gone (scripts/tests/memo-temporal.test.mjs:197-221).

scripts/lib/memo/eval/recall-ab.mjs is a deterministic A/B over synthetic fact chains, with arms for no links, explicit links, auto-detected links, entity boost and the embedder. Its test asserts the baseline's stale rate exceeds 0.4 and the explicit arm's is zero. It also asserts the automatic detector covers fewer than half the chains, labelled "Recorded as a measurement, not an aspiration" (scripts/tests/memo-ab-eval.test.mjs:43-71).

Two gaps matter. No test exercises MCP memory_write and asserts its claim status. No test asserts that candidate text stays out of unified search.

The project has no paper. A design report in the tree cites Agentic Context Management (arXiv:2607.21503, 23 July 2026) as an input and states its mechanisms were not reproduced (docs/reports/2026-07-28-context-lifecycle-competitor-analysis-and-plan.md:6, :48).

11. For Your Own Build

Steal

  • Put the status verdict in the write function, not the caller. Derive candidate or verified from the identity and scope inside the append, and ignore any status the payload claims. Then test it with a poisoned environment.
  • Make supersession an appended event with an ACL. Store only the forward link, derive invalidAt at read, let the earliest supersede win, and deny unknown targets at write time so a dangling id cannot later retire a fact in another space.
  • Record the denial, not only the grant. A governance receipt for every DENY, with the reason code and the content-safety verdict, makes a sealed gate auditable.
  • Assert the limits of your own detector. An eval that fails if automatic supersede detection starts claiming full coverage keeps a heuristic honest.

Avoid

  • A fallback identity that is the most trusted one. When the no-identity branch is verified as local-user, every new integration that forgets the identity argument inherits publish authority. Make the absent identity the least trusted.
  • A trust filter on one arm of a merged search. If transcripts and beliefs share a result list, the status gate has to hold on both, or the transcript arm returns what the belief arm withholds.
  • A review queue with no reviewer. Shipping the deny-all seam before the broker turns every automatic memory into a permanent orphan.
  • Agent guidance that describes intended behaviour. A skill that tells the model its writes are candidates, when they are not, trains it to trust a gate that is absent.

Fit

For a single developer who writes memos by hand and uses the one-shot wrapper little, this is a sound local store: append-only, repairable by hand, with real supersession and time travel, no model and no service. For anyone relying on the candidate gate to keep model-authored claims out of shared memory, it does not yet do that job. The MCP and Pi tools let the model publish verified facts under a human principal, and the only path that produces candidates leads to a queue no command can empty. Adopt the storage and supersession design; treat the governance as a specification awaiting its broker.

12. Open Questions

  • Is a broker planned that would supply hasBrokerReviewAuthority, and would it also sign the MCP server's writes?
  • Does a rebuilt ContextDB index pick up the legacy workspace-memory mirror, making superseded memo text searchable through the ContextDB arm?
  • How large does events.jsonl grow in practice before per-query full reads and scoring become noticeable?
  • Is observed produced anywhere, or is it reserved?

Appendix: File Index

  • Storage and schema: scripts/lib/memo/storage/events-write.mjs, events-read.mjs, normalizers.mjs, provenance.mjs, temporal.mjs, pinned.mjs, lock.mjs, paths.mjs; scripts/lib/aios/state-root.mjs; mcp-server/src/contextdb/sqlite/schema.ts, mcp-server/src/contextdb/core.ts.
  • Write path: scripts/lib/memo/cli/commands/events.mjs, scripts/lib/memo/autopilot.mjs, scripts/lib/memo/declaration.mjs, scripts/lib/memo/import-external.mjs, scripts/lib/memo/cli/legacy.mjs, scripts/lib/ctx-agent-core/run.mjs, scripts/lib/lifecycle/session-hooks/close.mjs.
  • Retrieval: scripts/lib/memo/storage/query.mjs, feedback.mjs, embedding.mjs; scripts/lib/search/unified-search.mjs.
  • Context assembly: scripts/lib/planning/turn-recall.mjs, scripts/lib/planning/user-prompt-submit.mjs.
  • Governance and background: scripts/lib/memo/storage/candidates.mjs, scripts/lib/lifecycle/dream/governance.mjs, index.mjs, scripts/lib/memo/autodream-auto.mjs.
  • MCP and integration: scripts/memory-mcp-server.mjs, scripts/lib/native/emitters/codex-config.mjs, scripts/lib/components/pi/mcp-adapter.mjs, packages/aios-pi/lib/tools.mjs, packages/aios-pi/lib/aios-cli.mjs, skill-sources/memo/SKILL.md.
  • Tests and evals: scripts/tests/memo-scope.test.mjs, memo-provenance.test.mjs, memo-temporal.test.mjs, memo-candidate-governance.test.mjs, memo-authority-env.test.mjs, memo-ab-eval.test.mjs, dream-governance.test.mjs; scripts/lib/memo/eval/recall-ab.mjs.

Recorded searches

Checked against the checkout at the pinned revision.

  • rg -n 'runtimeIdentity' -g '!scripts/tests/**' -g '!**/*.test.*' — callers passing an identity are autopilot.mjs, import-external.mjs, the candidate and dream CLIs and a benchmark; memory-mcp-server.mjs and cli/dispatch.mjs:400 pass none.
  • rg -n 'recordAutomaticMemory|promoteMemoryCandidate|hasBrokerReviewAuthority|trusted_authority_unavailable' -g '!docs/**' — one automatic-write caller (ctx-agent-core/run.mjs:424); both hasBrokerReviewAuthority definitions return false.
  • rg -n 'parseMemoryDeclaration\(' -g '!scripts/tests/**' scripts packages — ctx-agent-core/run.mjs:412 and planning/turn-recall.mjs:178 only.
  • rg -n 'memory_write|handleMemoryWrite' scripts/tests — no match; the Pi test names aios_memory_write only.
  • rg -n -i "forget|'delete'|'remove'|'rm'" scripts/lib/memo scripts/lib/cli/parse-args/memo.mjs scripts/memory-mcp-server.mjs — no match.
  • rg -n -i 'tombstone' -g '*.{mjs,ts,js}' . — dream proposal actions keyed on eventId and capability names in tests; no value-keyed record.
  • gh api repos/rexleimo/rex-harness/git/trees/869b87fee3283cbfcca61d86dbf537e44c5ba202?recursive=1 --jq '.tree[].path' | grep -i -E 'memo|memory|recall' — no match over the untruncated 217-entry tree.
  • grep -rliE 'arxiv|bibtex|@article|@misc|doi\.org' . --exclude-dir=.git --exclude-dir=node_modules — two design reports citing other work; no CITATION.cff.

History

2026-09-30 — 6e2910a9… — first reading, at the head of main, a commit dated 28 September 2026. Five marks: trust_state, bitemporal, scope_enforced, audit_log, negative_eval. Screened before reading: 1 auto-run surface (.gitmodules), no build-time execution, 14 dependency files inside the cooldown — every file in a depth-1 clone dates to the tip — and 6 unpinned surfaces; AGENTS.md, CLAUDE.md and GEMINI.md recorded as data. The rex-harness submodule was not cloned. Read with grep and sed; nothing installed, built or run. Only the memory subsystem is covered.