Personal Wiki
The durable output of an agent's experience is a human-readable wiki of skills and knowledge that the next agent — and the next person — can read and revise, not a vector store or a weight update; compiling experience into that form is the consolidation step the memory field left hand-rolled.
Someone ran it in their own harness. Artifact, no protocol. Decays fast.
Confidence
53%human-committedExpiry
22duntil review · 25 Sep 2026Lead time
—not yet mainstream · opened 3 Jul 2026Ownership
TOTom Okaforfortnightly cadenceWhere it is
The idea arrived as a paper pinned to the whiteboard and became a field when two engineers found they had built the same thing independently: a Markdown wiki the agent writes to at the end of a session and reads from at the start, with pages for skills ('how to run the eval harness'), facts ('the staging DB is read-only on Fridays') and mistakes. The WikiSkill paper formalised it — compile trajectories into skill pages, retrieve pages not chunks, let humans edit — and reported skill-reuse gains on long-horizon benchmarks. The lab's experiment (x-personal-wiki) is running: same tasks, same agent, with and without the wiki, measuring repeat-task time and error rate. The early signal is that the wiki helps on repeat tasks and that the compilation step produces confidently wrong pages at a rate nobody has measured, which is the same failure mode the memory field found in consolidation. The gate is tooling: nothing exists that compiles, versions, reviews and retires wiki pages as a unit.
Why a Quantium decision hinges on it
Every delivery pod using agents re-discovers the same environment facts and the same skills, engagement after engagement. A wiki that compounds across sessions and is readable by the humans who own the engagement is the artefact the firm can keep when the pod rotates — and the one it can hand to a client. It is also the honest middle path between agentic memory (which the firm has) and continuous learning (which it will not have for years): experience that persists without touching weights.
Field attributes
Position
What is demonstrated, what is hype, what would have to be true.
The shape every position request answers. Signal-tier fields carry a draft; assessed and tested fields carry a validated one.
- 01Two engineers independently built a session-end wiki compiler in Claude Code and both saw repeat-task time fall on their own work (tried, two instances, no control).
- 02WikiSkill reports 21–34% improvement in long-horizon benchmark success when agents retrieve compiled skill pages instead of raw trajectories (band-1 paper, not lab work).
- 03Interim from x-personal-wiki: repeat-task time down 44% with the wiki on a 30-task internal set; 6 of 52 compiled pages contained a factual error the agent later acted on.
- 01'Agents that learn from experience.' They compile it; they do not learn it. The distinction matters because a wrong page is inherited by every future session until a human notices.
- 02Vendor 'skill libraries' as a product. The value is in the compile-and-review loop, which is engagement-specific and will not ship in a SaaS.
- 03The claim that this makes agentic memory unnecessary. It is the consolidation half of memory with the output made readable; the episodic half is still needed.
- 01A measured error rate for compiled pages and a review loop that catches wrong pages before an agent acts on them — the current 6-in-52 is the number to beat.
- 02The wiki helping a different agent, on a different task, than the one that wrote it — transfer, not replay. The experiment measures replay.
- 03Tooling that treats a wiki page like code: versioned, reviewed, retired on staleness, with an owner.
- 01Let x-personal-wiki conclude; if the repeat-task result holds and the page-error rate is measurable, promote to tested and write a not-yet recommendation pending the review loop.
- 02Build the review loop as the lab's next Type 2: a second agent challenges every compiled page against the trajectory it came from.
- 03Trial the wiki as a handover artefact on one internal project rotation before any client engagement.
Signals · 10 in this cluster
What the cluster is made of.
Every item carries its source, tier and sightings. Detector-found signal sits beside human drops; downstream they are indistinguishable except by provenance.
WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Reuse
Compiles agent trajectories into a structured wiki of skill, fact and failure pages, retrieved at page granularity; humans can edit. Reports 21–34% higher success on long-horizon benchmarks against trajectory retrieval and a 2.1× reduction in repeat-task steps. Notes but does not measure compiled-page error rates. Pinned to the whiteboard; the field's origin.
extracted claimCompiled, page-granular skill knowledge outperforms trajectory retrieval on long-horizon agent tasks.
x-personal-wiki interim: repeat-task time −44%; 6 of 52 compiled pages wrong and acted on
Thirty internal tasks, same agent, with and without a session-compiled wiki. Repeat-task time fell sharply. A page audit found six factual errors, four of which the agent acted on in a later session. Experiment still running; interim only.
extracted claimThe wiki cuts repeat-task time and produces wrong pages agents trust; both effects are large.
Frontier lab hiring 'Skill Compilation Research Engineer'
Posting describes 'turning agent trajectories into reusable, verifiable skills'. Argus inference: the compile step, not just storage, is being built first-party. Carried as inference.
Foundation lab ships first-party 'skills' — reusable, file-based capability packages agents load on demand
A skills primitive with a file layout and loading semantics. Storage and retrieval become a platform feature; the compile-and-review loop is not included. Pricing tell: skills load at cached-token rates.
extracted claimSkill storage and retrieval are becoming a first-party platform primitive.
Logged from Claude Code: hand-corrected a wiki page the agent wrote; the next three sessions used the correction
Product engineer found a compiled page asserting a test fixture was safe to mutate; it was not. Edited the page in Markdown; subsequent sessions respected it. The readability is what made the fix possible. Tried tier.
'What does the pod leave behind when it leaves — apart from the code?'
Asked by a retail client's engineering lead at a rotation review. Logged by the retail sector owner. Not a request for a wiki; the question a wiki answers.
SkillBench-LH: long-horizon repeat-task suite with skill-reuse scoring
Public benchmark released alongside the WikiSkill preprint. The lab's 30-task set is adapted from it, with the repeat-task split kept and the transfer split added because the paper did not measure it.
'Your agent doesn't need memory. It needs a wiki.'
Argues a compiled wiki replaces episodic memory entirely. Contradicted by the memory bench: the wiki holds conclusions, not what happened, and audit and consistency tasks need the latter. Kept as the strongest overreach.
agent-wiki — session-end compiler that writes Markdown skill pages into a git repo
Open-source implementation predating the paper, with pages as git-versioned Markdown and a PR per session. The version-control choice is why humans edit it; the lab's experiment uses a fork.
'Skills are the new prompts' — three talks in one track
Demand-band signal: the word 'skills' in the sense of reusable agent capability appeared in eleven talk titles. Cross-band ignition with the August primitive release and the paper.
Claims · 4 supporting, 1 refuting
The atoms.
A document cannot go stale; an assertion can. Claims are immutable and stamped with the extractor that produced them, so staleness, diffs and the graveyard operate at claim level.
Retrieving compiled skill pages beats retrieving raw trajectories on long-horizon repeat tasks by a wide margin.
The compilation step produces confidently wrong pages at a material rate, and agents act on them.
Foundation labs are converging on a first-party 'skills' primitive, which will commoditise storage and leave compile-and-review as the differentiator.
Human-readable wiki pages are edited by humans in practice; the readability is used, not decorative.
A compiled wiki makes an explicit episodic memory store unnecessary.
Position history · the diff is the product
3 validation runs against a fixed brief. Confidence 45% → 53%.
Experiment interim: repeat-task time down 44%; 6 of 52 compiled pages wrong and acted on. The page-error rate is now the primary open question and the gate is tooling for the review loop.
- The compilation step produces confidently wrong pages at a material rate, and agents act on them.
- c-personal-wiki-1 ↑ 0.6 → 0.7
Scoring · ordinal bands
Agents propose. A named human commits.
Uncommitted scores are visibly marked and never leave the building. Bands, not point estimates — false precision is the tell that a number was generated rather than derived.
Impact
committed · AWCompounds across every agent engagement the firm runs; bounded by the page-error problem until the review loop exists.
Timeline
committed · TOUsable on internal work now; client handover after the experiment concludes.
Cost
committed · TOOne engineer, the running experiment, and a Type 2 for the review loop.
TAM
agent-estimatedAgent-estimated from agent-platform 'skills' tooling spend. Small as a market; the value is internal. Uncommitted.
Cost of being wrong
agent-estimatedA wrong page inherited across sessions is a slow, compounding error rather than an incident. Agent-estimated.
Demand
committed · DSNo client has asked for it by name; one retail client asked what the pod leaves behind when it leaves, which is the same question.
Workforce readiness
committed · TOEngineers adopt it unprompted; the review discipline is the part nobody does yet.
Relevance · per vertical
Why it matters here, or explicitly does not.
Ranking is per vertical, not global. Sector owners commit notes against agent drafts.
Every pod rotation loses environment knowledge and skills. A compiled, readable wiki is the first handover artefact an agent can produce.
Mechanism · Wiki compiled per engagement, reviewed by the pod lead, handed to the client with the code.
A wiki that captures how the agent handled a client's system is also a record that may contain client data and operational detail. Handover is attractive; retention is a governance question.
Mechanism · Would need a page-classification step and a retention policy before pages leave the engagement.
The merchandising assistant re-derives category doctrine each session; a compiled doctrine wiki is the fix the memory field pointed at.
Mechanism · Semantic-store consolidation from agentic-memory written as reviewable pages instead of opaque summaries.
Red team · the strongest case against
The strongest case against: this is documentation, and documentation rots. The lab has rediscovered that writing things down helps and that written-down things go wrong, and has attached an agent to both halves. The 6-in-52 error rate is the whole story — a page an agent wrote and an agent trusts is a hallucination with a URL, and the review loop that fixes it is the human effort the wiki was meant to save.
- —The interim result measures replay: the same agent on the same tasks it wrote pages about. Transfer to a different task or agent is unmeasured and is the only version worth paying for.
- —Wrong pages compound. An episodic store holds what happened; a compiled page holds what the agent concluded, and the conclusion inherits every error in the trajectory it came from.
- —Foundation labs are shipping the primitive. In twelve months the storage and retrieval are free, and what remains is a review discipline the firm has never sustained for its human-written wikis either.
Source diversity
- ML / agents research30%
- Open-source tooling20%
- Foundation lab20%
- Internal / Engel30%
A field supported by one epistemic community is a flag, not a finding.
Cross-pollination · typed joins
Connected, not merely similar.
Enabling, compounding, substituting, blocking. A satisfied dependency trigger is a far stronger signal than semantic proximity.
The wiki is the consolidation step of the memory layer with a human-readable output; both need the episodic store underneath.
If weights can learn from use, compiled pages become unnecessary. Until then the wiki is the substitute — experience that persists without touching weights.
A learning agent's decision memory needs a reviewable form; compiled pages are one.
The lab's own harness knowledge is the first wiki, and the one the experiment runs on.
Share graph
Provenance running forward.
Discovery, not accountability. No counts, no rankings, no rollups to managers.
Convergence · who else is here
- AWAdam Witanowski · Lab Director (acting)2 drops
- OGOllie Grant · Product engineer2 drops
- MTMei Tanaka · Research lead · evals2 drops
- TOTom Okafor · Research engineer · agents1 drop
- ?Anonymous · Anonymous drop1 drop
- DSDev Sharma · Sector owner · Retail1 drop
Several people’s drops meet here. An informal working group already exists and probably does not know it.
Lineage
What this field produced, and what it killed.
Experiments, recommendations and graveyard entries stay attached. The reasoning that killed a claim is the reusable asset.
Open questions · return to the pile
Every run leaves a record. Separately, its question either closes or returns to the pile with notes — which is what the next person proposing the same thing will see.
- 01What is the compiled-page error rate on a full run, and does a challenge agent bring it under 2%?
- 02Does the wiki help a different agent on a different task — transfer — or only the agent that wrote it on replay?
- 03What has to be stripped from an engagement wiki before it can be handed to a client or retained by the firm?