Back to Agents That Account for Themselves
working paper2026KARL paper

Trajectory Memory Ledger

The Trajectory Memory Ledger is the formal paper behind KARL. It describes the append-only schema, six-signal reward engine, OAPL-Lite export, and entity bridge used to turn real agent work into reviewed training data.

Paper workspace

Live draft structure

working-draft

Artifacts

Markdown paper source

The Trajectory Memory Ledger paper is mapped from the KARL paper source.

source-only

Editable source

Paper source exists. Public artifact should be rendered after privacy-sensitive example review.

Source anchors

karl/paper/karl-paper.md

Trajectory Memory Ledger deployment notes in MEMORY.md

KARL trajectory-ledgerd and SFT exporter

Method tags

ledger schematool-use recordingreward scoringSFT export

Ingest intersections

trajectory-ledgerexperience-replayschemareward

Status

Paper drafted from the KARL deployment corpus; public abstract safe, raw traces private.

Key claims

01

Experience replay for coding agents should preserve full tool-use trajectories.

02

Reward can be derived from observable behavior without explicit preference labels.

03

Schema normalization is what makes cross-project training data possible.

Public reading note

Readable public page; full manuscript needs privacy review.

Standard skeleton

What this paper must keep proving

Schema

problem

A coding agent cannot improve from work it cannot replay, score, or normalize across projects.

method

Record complete tool-use sessions into an append-only schema and score them with a six-signal reward engine.

implementation

Four-tap instrumentation, Rust ledger daemon, schema-v2 normalization, SFT export, and entity bridge.

data

Real private session traces; public release should expose schema and aggregate behavior, not raw conversations.

evaluation

Schema consistency, reward separation, export quality, and downstream training/routing lift.

references

Agent observability, experience replay, process reward models, telemetry schemas.

openQuestions

Which anonymized examples can be safely published as canonical ledger rows.

Checkpoints and references

Proof chain

paperpending

Claim checkpoint

central-claim slot

Every central claim must point to a proof anchor or remain labeled as speculative.

implementationpending

Implementation checkpoint

implementation-map slot

Every method should identify the code path, harness, schema, or protocol that embodies it.

experimentpending

Evidence checkpoint

evidence-manifest slot

Every reported result should point to run IDs, packet IDs, data snapshots, commits, or review artifacts.

external-referencepending

Reference checkpoint

references slot

Every external claim should resolve to a cited paper, benchmark, standard, or documented prior system.

paperpending

Release checkpoint

release-gate slot

Every PDF needs a named condition before it can move from draft to citation-ready.