Where the 2.6 line stands¶
Updated: 2026-10-01. This page says how far the 2.6 line has come, item by item, with the evidence for each. The line's page says what the line builds and what "done" means; the release notes say what each release contains. Where this page and a release disagree, the release is right.
2.6.0 is the line's first final release. 2.6.1 followed the same day, 2.6.2 the
next (2026-10-01) and 2.6.3 the day after (2026-10-02), and a plain install
resolves to 2.6.3. 2.6.4a1 (2026-10-02) and 2.6.4a2 (2026-10-03) are
pre-releases, installed only by exact version.
Its pre-releases are on PyPI as 2.6.0aN and 2.6.0bN; the 2.5 line is now
Candidate (SUPPORT.md).
The four states¶
Every row below carries exactly one of these. They are kept apart so that a design is never read as a feature.
| State | Meaning |
|---|---|
| Released | In a published release. Anyone can install it and call it. |
| In development | Code exists on master, but it is unreleased, or released without the measurement that decides its default. |
| In research | A hypothesis, an experiment or a design comparison. No server code yet. |
| Withdrawn or reworked | Measured and rejected, or found wrong and being redesigned. These rows stay on the page. |
Against the line's completion conditions¶
The numbering follows what "done" means.
| # | Condition | State | Evidence |
|---|---|---|---|
| 1 | The recall process and Cued Recall ship behind a gate | Released in 2.6.0a8, v0.3 | The loop's basic form: a caller may pass time_cue (when, with a confidence); the server also searches that period to a depth of its own, moves a row found there up by at most 3 / 2 / 1 places after the quality gate, holds as many seats for records that search found which the answer does not hold (one the count cut included), and widens the period once if it holds nothing (#325, #331–#333, #340, design). Opt-in per call. Measured under a pre-registered rule on the LMEB TMD task (1,167 questions that state a time): mean NDCG over the returned rows rose from 0.189 to 0.281 with the cue and to 0.204 with as many extra uncued rows (results). On LongMemEval the previous policy was not distinguishable from no cue (results). Cue propagation and richer stopping are later. |
| 2 | The final re-sort has been decided | Released in 2.6.0a7 | With confidence enabled, the confidence score no longer re-sorts or gates recall, and CPERSONA_CONFIDENCE_ORDERING=legacy restores it. The far weight and an age weight are one prior function that orders only what the gate admitted, both at identity defaults until measured (#322, design). The far weight has since been measured under a pre-registered rule: at a reach of 200,000, no constant weight kept the recent answers while keeping what the far list buys, so its default stays at 1 (results). |
| 3 | Depth and count are separated | Released in 2.6.0a2 | #274. CPERSONA_RECALL_DEPTH_FLOOR defaults to 0, which keeps the 2.5 coupling until a depth is chosen by measurement. |
| 4 | Reconstructive Recall exists as a tool | Released in 2.6.0a2, extended in a3 | The tool: #274. Breadth before depth and the payload budget: #280, #286. One item shape, and a response that says what it dropped: #288, #289, #290. Measured so far: a count replay over ceilings 1 to 10, which by its own statement does not choose a default; and a pre-registered reader study whose first run found no reduction in what a reader takes in, and whose re-measurement after two changes met the registered rule on 18 questions. The default count is still a contract choice, not a measured optimum. |
| 5 | Adaptive fusion beats the raw embedding on both models | In research | The design and the stage-by-stage loss analysis it rests on: #260, #261, design record. No server code. |
| 6 | Every benchmark failure has a code and a replayable trace | In development | Released in 2.6.0a7: on request, recall and reconstruct return a trace of what each stage kept, dropped and reordered, and a tool in benchmarks/ assigns a confirmed failure code from a trace and the answer's references (#324, design). Not yet applied to every benchmark failure. |
| 7 | The accuracy–token–latency–memory frontier has moved | In research | One axis measured. On the 15 LMEB tasks whose haystack can be one caller's memory, at ten rows, 2.6.0 in its best configuration ranks the relevant memory higher than 2.5.12 in its best: macro NDCG@10 57.36 → 59.12, 14 tasks up and 1 down, under a pre-registered rule (results). Tokens, latency and memory are not measured, so the frontier is not yet shown. |
| 8 | The open quality debt of the 2.5 baseline is closed or carried with reasons | In development | The calibration a benchmark run measured under is now recorded with the run: #258. The other items are open. |
Released in this line, beyond the conditions¶
| Release | What it added | Evidence |
|---|---|---|
| 2.6.0a1 | Full-text normalisation that keeps identifiers searchable | #271 |
| 2.6.0a3 | Long records divided into nodes, kept out of the search index; a reconstruct quote taken from the best node; part of a record expanded through get_contents |
#282–#287, design record |
| 2.6.0a4 | Declared associative memory: entities, aliases and relations the agent states; reconstruct reads them as cues with a bounded walk; traverse returns a declared neighbourhood |
#291–#295, design record |
| 2.6.0a7 | The episode-boundary penalty off by default: it was found to push the record holding an answer down the ranking on a store where each session ends in an episode summary | #320 |
| 2.6.0a8 | An opt-in propagation seat: one held place for the row a deeper ranking of the same recall leads to from the answer's first row (CPERSONA_RECALL_PROPAGATION_SEAT, off by default) |
#338 |
| 2.6.0a8 | Recall and reconstruct call each stage through a provider slot, and a traced recall records the budget it declared and what each stage received; nothing a caller receives changes | #335, #336 |
| 2.6.0b1 | No feature: fixes for the eleven MEDIUM defects a review of 2.6.0a1..a8 found, among them a project's alias other projects could read, a qualifier on the next line left out of a quote, reconstruct walking with no bound on its hops, an episode quote measured 10 characters off, and node sets the reader rejected that nothing rebuilt |
#356, #357, #358, #359, #361 |
| 2.6.0b2 | No feature: fixes for the five remaining MEDIUM defects, among them an in-memory export whose header could disagree with its body, an index status that reported rows each query reads as none, a list cut by its row cap that looked complete, a session_key with no length bound, and an index path that held a copied window whole |
#363, #364, #365, #366 |
| 2.6.0 | The line's first final release, and no feature: block reach on by default, reconstruct at 10 items when count is omitted (together, the configuration the release recommends for answering from memory), and fixes for forty-three LOW defects the alpha review registered, among them block sets that crossed a node end, a time cue past the representable range that raised, a failing build that held every queued task, and held rows that earned a recall count no gate granted. |
|
| 2.6.1 | Fixes and additions from a production report: the contiguous index can be rebuilt on Windows while a server holds it (bug-503); what a recall score means, and why it cannot say that nothing matches (contract §12); an example Claude Code hook that prompts recall (runbook); get_session_findings names the SuperAuditor version it conforms to (1.1) |
#377, #378, #379, #380 |
| 2.6.2 | The binary coarse search, off by default (design): a one-bit index of every record beside the contiguous index; far seats, two held places for records past the scan window (CPERSONA_FAR_SEATS_ENABLED); the rest of a time cue's period searched past the vector half's cap (CPERSONA_CUE_COARSE_ENABLED); the number of candidates set at 256 by a registered measurement (results). Also a coverage ledger in the recall trace, an empty-query time cue that lists a period's records by their own time (bug-502), and PyJWT 2.14.0 or later |
#382, #383, #384, #387, #389, #390, #391, #392 |
| 2.6.3a1 | Pre-release. Every stored vector is labelled with the model that produced it (schema 18): the fingerprint a backend reports from GET /capabilities (CEmbedding 0.9.0 and later), else the model the api transport sends or the name configured. check_health reports the vectors a different model wrote (embedding_model), and CPERSONA_EMBEDDING_MODEL_MODE=reject stops comparing them with the query; the default warn changes nothing recall returns (configuration) |
#395 |
| 2.6.3 | The final of 2.6.3a1, with no change to the code: a plain install now labels every stored vector with its model and migrates the schema to 18. It was cut after eight hours of production use on 2.6.3a1 with no restart, no error and every new vector labelled | |
| 2.6.4a1 | Pre-release. Reconstruct v1.2 (results): blocks are ranked by their stored int8 vectors, passages that touch are quoted as one, the head quote is sized by the item's place (CPERSONA_RECONSTRUCT_QUOTE_CHARS, 800, for the first CPERSONA_RECONSTRUCT_FULL_QUOTES, 5, items and CPERSONA_RECONSTRUCT_TAIL_QUOTE_CHARS, 400, after them), and the default response leaves to the trace what a reader does not act on. On the private real-use pack's 150 test questions the evidence quotes shown rose from 133 to 146 of 196 and the first response's median size fell to 0.662 of 2.6.3's. Also check_health gains coarse_index: while a coarse-search setting reads the index, a missing or unusable one is reported, and fix=true builds it |
#398, #399 |
| 2.6.4a2 | Pre-release. A time cue searches the part of its period past the scan window through the coarse index by default (CPERSONA_CUE_COARSE_ENABLED unset): with no usable index that part is not searched, the recall is the one the setting off gives, and time_cue.remainder says so; true keeps the 2.6.2 meaning (binary coarse search). Nothing builds the index: a recall whose scope holds 1,000 or more records past the window with no index carries a suggestion once per session for the agent to relay, and check_health reports the missing index as info while only this mode reads it |
#405 |
The associative layer is released and off by default. Whether it becomes the default is decided by its own A/B run; no result of that run is recorded yet.
Withdrawn or reworked¶
| What | What happened | Evidence |
|---|---|---|
| The reference panel for adaptive fusion, as first specified | It identified no null: the two densities it specified were the same distribution, so the evidence it defined could not separate a mixture from its own null. The definition was replaced before any code was built on it. | #262 |
rsf on a fixed channel scale (the bug-247 fix) |
It passed its pre-registered rule on LongMemEval at ten rows, where it also returned more rows. On LMEB Track B it scored lower on 19 of 22 tasks (macro 58.53 → 57.51), so it was reverted before release. The gate defect it addressed is open again. | results |
rsf gated on a fixed scale with its order unchanged (the second bug-247 design) |
It passed both of its pre-registered rules: no LMEB Track B task fell (macro 57.96 → 57.97) and no LongMemEval question fell at ten rows. It was not merged: under the calibrated fused gate a deployment runs by default, it chose a threshold that emptied or thinned most answers on a private pack of real agent memories (rows at ten 9.89 → 2.16, evidence reached 295 → 204 of 389). bug-247 stays open. | results |
What is not on this page¶
Dates. The roadmap is descriptive and this page follows it: a row moves when its evidence exists, not when a date arrives.