Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

Scope personal coding memory to a bounded MVP (#678)

Deefa8f966c781b
diff --git a/docs/plans/2026-09-30-coding-memory-review-log.md b/docs/plans/2026-09-30-coding-memory-review-log.md
index 48c375757..fa1b0ec0c 100644
--- a/docs/plans/2026-09-30-coding-memory-review-log.md
+++ b/docs/plans/2026-09-30-coding-memory-review-log.md
@@ -1,5 +1,12 @@
# Coding memory: sequential review and implementation
+Current priority: the user replaced open-ended research with a
+[bounded personal-memory MVP](2026-10-03-personal-memory-mvp.md) on October 3.
+Council and benchmark work is deferred until the user reviews the MVP and
+authorizes further scope. The
+historical entries below record evidence and remaining broader work; they do not
+expand the current milestone.
+
Status: in progress, 2026-09-30. The objective remains a thoroughly reviewed, useful coding memory system across agent frameworks. Passing an isolated core suite does not establish completion or perfection.
These reviews are performed by Codex using the published principles collected in the [historical council](2026-09-30-coding-memory-council.md) and [modern practitioner study](../research/2026-09-30-modern-coding-memory.md). They are not personal participation, simulated quotations, or endorsements by the named programmers. Each perspective produces a concrete question, a change or open requirement, and evidence needed to close it. Review again after integration, not just after the design document.
diff --git a/docs/plans/2026-09-30-tim-memory.md b/docs/plans/2026-09-30-tim-memory.md
index 3f485ebb6..1c3738b09 100644
--- a/docs/plans/2026-09-30-tim-memory.md
+++ b/docs/plans/2026-09-30-tim-memory.md
@@ -1,5 +1,11 @@
# Coding memory across agent frameworks
+Current delivery milestone: the user explicitly scoped a
+[two-hour personal-memory MVP](2026-10-03-personal-memory-mvp.md) on October 3.
+The broader research and rollout gates below remain a later roadmap, rather than
+prerequisites for that bounded review milestone. Its evidence and consent rules
+still apply; deferred checks must not be described as completed.
+
Status: experimental implementation available for review, 2026-10-01; quality validation remains incomplete. The memory core, native transcript reader, durable learning loop, restricted inference adapters, worker isolation and owner library have executable tests. Opt-in host capture, scoped collection recall, native prompt adapters and the desktop Memories viewer are connected. Synthetic native probes cover next-user-prompt delivery after resume, manual compaction and a model change in Claude 2.1.286/2.1.287 and trusted interactive Codex 0.159.3. The older Codex 0.159.0 prompt certificate remains. Local synthetic recall performance has been measured; real extraction quality, full native lifecycles and production integration remain unverified. The [sequential review log](2026-09-30-coding-memory-review-log.md) records findings, fixes, and remaining completion evidence.
Codex **0.159.3 prompt delivery and background extraction have separate status**. Its trusted prompt
diff --git a/docs/plans/2026-10-03-personal-memory-mvp.md b/docs/plans/2026-10-03-personal-memory-mvp.md
new file mode 100644
index 000000000..6d71f48cf
--- /dev/null
+++ b/docs/plans/2026-10-03-personal-memory-mvp.md
@@ -0,0 +1,99 @@
+# Personal coding memory MVP
+
+Status: bounded review handoff prepared; live acceptance remains open. Scoped by
+the user on October 3, 2026. This is the current delivery
+milestone. The broader council, architecture and benchmark plans remain a roadmap;
+their complete research programs are not prerequisites for this review milestone.
+
+The finish line is a reviewable experience that demonstrates useful memory about
+how this user builds Harness, with concrete evidence. It is not a claim of perfect
+or generally validated memory. The user decides on further investment after review.
+
+## Scope and acceptance
+
+One user, one machine, the Harness coding project, and the existing Companions DSH.
+Keep the viewer left and the actual agent terminal right. Use the existing
+experimental setting and the model selected for that DSH. Memory belongs to the
+Harness profile/project and remains independent of avatar and agent framework.
+
+1. **Learn something real.** Produce three to five accurate, actionable memories
+ from already consented real conversation evidence. Every displayed assertion
+ needs an inspectable source quote and appropriate project/condition boundaries.
+ Do not fill a quota with speculation or turn a suggestion into a settled decision.
+2. **Continue learning.** Demonstrate one new eligible interaction moving through
+ capture, extraction and storage. A history import or hand-seeded record alone
+ does not establish automatic learning. Show a truthful waiting/error state if
+ the selected model cannot run.
+3. **Help on a later task.** Demonstrate a fresh coding session receiving a relevant
+ memory and using it correctly, without repeating the preference in that task.
+ Check one unrelated request does not receive that memory. Keep the shared
+ storage/recall contract across existing framework adapters; exhaustive native
+ version and lifecycle certification is deferred.
+4. **Make it inspectable and editable.** Show memories and evidence in the existing
+ Memories viewer. Verify a correction affects later recall and a forgotten memory
+ is absent from later recall. Do not claim old conversation context was erased.
+5. **Hand over a reviewable result.** Provide the exact build/source revision, a
+ short walkthrough, and a pass/fail account of these behaviors. Distinguish
+ delivery, correct use and user confirmation. Do not present synthetic fixtures
+ as learning from real sessions.
+
+## Time box and work order
+
+Start: 2026-10-03 05:27:27 UTC. Handoff target: 07:27:27 UTC, two hours later.
+The cap includes implementation, affected verification and review preparation.
+
+- First 15 minutes: freeze scope and identify the shortest working path through
+ the existing code; resolve or name the live-model connection gap.
+- Next 45 minutes: establish the real memories and one new-learning example,
+ making only changes needed for this acceptance list.
+- Next 40 minutes: verify fresh-session use, exclusion, correction and forgetting,
+ and prepare the review experience.
+- Final 20 minutes: review, complete necessary affected checks, and deliver the
+ build/walkthrough with every remaining gap explicitly identified.
+
+At the cap, report the concrete result and every unmet criterion. Do not hide an
+unavailable selected model behind another research experiment or call the MVP
+complete without evidence. Further scope requires a user decision.
+
+## Deferred work
+
+No new model/classifier comparisons, literature rounds, exhaustive benchmark
+matrix, framework integrations, memory taxonomy, multi-machine synchronization,
+firmware work or feature expansion. Existing privacy, source attribution, project
+isolation and user controls remain required. The broader
+[architecture](2026-09-30-tim-memory.md) and
+[review program](2026-09-30-coding-memory-review-log.md) remain a later roadmap.
+
+## Starting evidence and constraints
+
+The core service, framework adapters, experimental setting, Memories viewer and
+companion-terminal recovery action are merged. The last read-only check of the
+installed daemon reported `waiting_for_model` and capture paused with
+`memory_backlog_full`. Merged code is not necessarily installed code.
+
+Personal evidence and memory artifacts stay local. Existing consent boundaries
+apply; rescoping does not authorize sending private conversations to an unapproved
+provider or silently replacing the selected companion model. Local service
+installation, opening the new review app and restarting the companion were
+subsequently authorized for the live test. Preserve saved conversations and the
+existing review app. No release, installed app replacement or firmware deployment
+is included. The unfinished synthetic comparison was stopped and archived locally;
+it is not completion evidence for this MVP.
+
+## Handoff boundary
+
+Normal desktop and matching CLI review artifacts have been built from the merged
+application source. A bounded local demonstration and the existing Memories UI
+have been exercised against an isolated review store. Personal excerpts, records,
+screenshots, model outputs and detailed findings remain local, outside Git.
+
+These checks do not establish new automatic learning in the installed application.
+The live capture, selected-companion extraction and storage criterion must still
+be demonstrated. Opening the normal desktop build alone does not update its
+installed daemon. The approved live test must use the selected companion model
+and stop if its provider refuses. The local review artifacts alone do not establish
+that this live path works.
+
+After the handoff, the next work is only the remaining live acceptance check when
+authorized. Do not resume the broader research program, add new experiments, or
+declare the MVP complete while that criterion remains unverified.