Threadnote

Local-first engineering memory and large-scale polyglot graph search for Codex, Claude, Cursor, Copilot, and the next agent you try.

[Threadnote @threadnoteio]

Your coding agent has the memory of a goldfish with sudo.

Threadnote gives it the decisions, handoffs, and code evidence that survive the chat.

Open source: https://t.co/5dAqGgdYMr

[Threadnote @threadnoteio]

8/8

Also in 5.2.0: managed OpenCode + Zeno setup (Zeno on macOS), resumable health checks and clearer Manager graph/operation status.

Start with one agent and one real task:
https://t.co/CA3ZxicF3q

Full release notes:
https://t.co/mooTHT1SNV

[Threadnote @threadnoteio]

1/8

Threadnote 5.2.0 is out.

Your coding agent remembers two production timeouts. Both confidently. Naturally.

Now Manager lets you compare conflicting memories and choose what stays current. The newest timestamp doesn't win by default.

A quick tour:

[Threadnote @threadnoteio]

7/8

Your note survived. Its source moved on.

5.2.0 can retain the reviewed source passage and report current-source status separately. Obsidian notes included.

Remembering the words isn't the same as checking whether they still apply.

[Threadnote @threadnoteio]

@starkbotha I don't have measured data for an Opus sweet spot. I'd replay the same task from one checkpoint at two thresholds, with fixed checks for missed constraints and patch correctness. Did 500k hold up across several tasks?

[Threadnote @threadnoteio]

@Devon8zzh For the skipped skills, I'd test one small step in a fresh chat: same instructions, same skill file. Check whether it reads the file before acting. Did the misses start after compaction, or from the start?

[Threadnote @threadnoteio]

@AlexChhk I'd make expiry a review trigger, not automatic deletion. Keep the why and the change that invalidates it: dependency upgrade, config change, bug fix. A stale decision can still explain today's code.

[Threadnote @threadnoteio]

@AlexChhk I'd add a 'revisit when' line to those decisions. Otherwise last week's workaround gets promoted to company policy. Do you track the reason, or just the final choice?

[Threadnote @threadnoteio]

@stevegraham I'd include failed approaches and their evidence in that handoff too. Otherwise the fresh chat gets a fresh opportunity to rediscover the same dead end.

[Threadnote @threadnoteio]

@stevegraham I'd give the new chat the goal, current diff and acceptance checks before the transcript. Fresh window, same baggage gets expensive. Does yours get a short handoff or the full history?

[Threadnote @threadnoteio]

The headline counts all measured work, including the failed attempt, per verified completion. Across the five assigned workflows, total tokens fell 57.0% and total time 32.7%. The article has both sets of numbers.

[Threadnote @threadnoteio]

65.6% fewer tokens.
46.1% less time.
Both per verified completion, including work spent on the failed attempt.

Your agent figured out the bug.
The next session would like to start from scratch. How thoughtful.

Threadnote 5.1.0 is out. The next session gets a handoff and code context it can use.

Our five-task continuation study:

5/5 patches passed checks, versus 4/5 with files alone.

One model. Same starting checkpoints. Threadnote handoff + code-graph query versus files-only continuation with no handoff. We haven't tested a standardized plain-text handoff yet.

Full results, timings and limits:
https://t.co/mQ1gR7ugKO

[Threadnote @threadnoteio]

@tinprofile Is most of that minute launching Codex, or rebuilding the task context? I'd time those separately before optimizing the handoff. Otherwise you can write a beautifully concise brief for a process that's still booting.

[Threadnote @threadnoteio]

Try it: https://t.co/5dAqGgdqWT

[Threadnote @threadnoteio]

New chat. Same repo. Apparently we’ve never met.

Threadnote keeps the useful bits between tasks, with sources you can check.

A 70-second tour of how that works.

[Threadnote @threadnoteio]

@RishwiBTweets I'd make scrolling an acceptance check: open a long page at a small viewport and have the agent reach the last control. 'Page rendered' is a very generous definition of done.