Upgrading from 2.5 to 2.6¶

This page takes an existing 2.5.x store to 2.6.0 in one pass. Each 2.6 pre-release documented only its own step in its release notes; this page puts the steps from 2.5.12 onward in one place.

2.6.0 is the 2.6 line's first final release (Current in the support policy), and the 2.5 line becomes Candidate: no channel serves it, and it stays reachable by exact version.

Before you start¶

Back up the database and its calibration file. Follow Backup and restore: an online sqlite3 "$CPERSONA_DB_PATH" ".backup 'cpersona-backup.db'", plus a copy of <CPERSONA_DB_PATH>.calibration.json, which lives outside the database. This backup is the only way back to 2.5 (see Going back to 2.5).
Check the embedding server. Building the overflow nodes for records already stored needs a server that reports token counts: CEmbedding 0.8.0 or later. An older server answers count: null, which is not zero.
Install 2.6.0. A plain upgrade now resolves to it; pin the version so you know which one you are running:

pip install 'cpersona==2.6.0'

What the first start does¶

On its first start, 2.6 migrates the database from schema version 13 (every 2.5.x release) to schema version 18 (17 for 2.6.0 to 2.6.2), one step at a time. No stored row is rewritten: each step adds tables and triggers, or one column. A step that fails is not recorded as done, so it is retried on the next start.

Schema version	Added in	What it adds	Anything to build afterwards?
14	2.6.0a3	`record_nodes`: the overflow tree, pieces of a record past the embedding window	Yes: nodes for long records already stored (below)
15	2.6.0a4	`entities`, `entity_aliases`, `entity_mentions`, `relations`: declared associations	No
16	2.6.0a5	`record_blocks`: block reach	Built for you: blocks for the records already stored, unless you turn block reach off (below)
17	2.6.0a6	`record_block_vectors`: one vector per block	Built with the blocks
18	2.6.3a1	`embedding_model` on `memories` and `episodes`: the label of the model that produced each vector. Rows already stored take an empty label, which means unknown	No: a vector is labelled when it is next written

2.6.0a1, 2.6.0a2, 2.6.0a7, 2.6.0a8, 2.6.0b1, 2.6.0b2, 2.6.0, 2.6.1, 2.6.2, 2.6.4a1 and 2.6.4a2 changed no schema.

After the first start¶

The gate is recalibrated¶

2.6.0a7 changed the scoring version, so a calibration stored by an earlier version is treated as stale. With the default CPERSONA_CALIBRATE_ON_MODEL_CHANGE=true, the server recalibrates the global threshold at startup. If both that and CPERSONA_AUTO_CALIBRATE are off, the stale gate is not applied and deep_check reports stale_scoring_version until calibrate_threshold is run.

Per-agent calibrations are not redone. A threshold or gate calibrated for one agent with calibrate_threshold is discarded with the old scoring version; that agent falls back to the global threshold and the heuristic gate until you run calibrate_threshold for it again. Preferences set with set_recall_precision are kept.

Build the overflow nodes¶

Long records stored before the upgrade have no nodes until they are built. Until then they are quoted from their start, and node_unavailable: no_nodes says so. Build them with the health check, which handles 50 records per run; repeat it until it reports none left:

check_health(agent_id="<id>", fix=true, checks=["missing_nodes"])

The repair never modifies a record, so it also covers locked memories. Run it again after changing the embedding model.

Block reach is on by default¶

From 2.6.0, block reach is on unless you turn it off. It makes text past a long record's embedding window reachable by search, and reconstruct quotes the block that matched (design).

The first start begins a bounded backfill of the records already stored. It embeds every block through the same embedding server that embeds your records, one call per batch of blocks; on this project's own store, 4,478 records divided into 70,130 blocks. check_health shows the progress as missing_blocks and moves it along under fix=true. Until a record's blocks exist, recall reaches it through the other arms, as it does with block reach off.
Each block is stored twice, as one bit per dimension and as one byte per dimension: 1,152 bytes of vector per block with a 1,024-dimension model (128 + 1,024).
Recall reads the index on every query that has a vector, and may return up to 2 rows beyond limit for records only the block arm reached. It has no effect where vector search is remote.
Recall gets slower and larger. On that store, on one machine, the median recall took 138 ms with the block arm against 17 ms for 2.5.12 (the query's embedding excluded), and the process peaked at 180 MB against 106 MB. The cost grows with the number of blocks; a store ten times larger has not been measured.
To turn it off, set CPERSONA_BLOCK_BUILD_ENABLED=false: no embedding calls, no rows and no queued work, and the reader follows it off. CPERSONA_BLOCK_RETRIEVAL_ENABLED=false alone keeps the index built and unread. Setting the reader on with construction off is a startup error.

If you ran 2.6.0a5 with block construction on, its block sets have no vectors and are rebuilt by the same backfill.

Behaviour that changed¶

Check these against what your deployment relies on. Each is off, or equal to 2.5, unless noted.

Block reach is on by default (2.6.0). A store queues the record's blocks and says so (blocks: {"status": "queued"}), recall can return up to 2 reserved rows beyond limit, and reconstruct quotes the block that matched. Above is how to turn it off.
reconstruct returns 10 items when count is omitted (2.6.0; it returned 1). With block reach on, this is the configuration 2.6.0 recommends for answering from memory. A deployment that relied on one item sets CPERSONA_RECONSTRUCT_DEFAULT_COUNT=1; lowering CPERSONA_RECONSTRUCT_MAX_COUNT alone lowers the default with it.
limit is the number of rows returned (2.6.0a2). How deep fusion looks is max(limit, CPERSONA_RECALL_DEPTH_FLOOR); the floor defaults to 0, so the ranking is the one 2.5 gave.
reconstruct needs the same per-agent read grant as recall (2.6.0a2). This matters only where access control is configured.
Confidence no longer re-sorts or gates recall (2.6.0a7). With CPERSONA_CONFIDENCE_ENABLED=true, each row still carries a confidence value, but the order is the fusion order. The profile row (update_profile) used to rise to the top under the old re-sort; it now comes last and a full limit can cut it. If something must reach the agent every time, put it in the client's instructions file or system prompt. CPERSONA_CONFIDENCE_ORDERING=legacy restores the old order.
The episode-boundary penalty is off by default (2.6.0a7). Deployments that used it to damp older sessions set CPERSONA_EPISODE_PENALTY_ENABLED=true.
A time cue can add up to three rows (2.6.0a8). With time_cue, recall returns up to limit + 3 rows (3 / 2 / 1 by the cue's confidence), and a cue that points only at the last 24 hours is not used (time_cue.ignored). Without a cue nothing changes.
declare_associations is marked destructive (2.6.0b1). Its retract argument deletes, so a client that asks before running destructive tools now asks before this one.
reconstruct walks at most 5 hops, as traverse does (2.6.0b1). A larger max_hops is lowered, and bounds.max_hops states the value applied.
An alias follows the scope it is declared in (2.6.0b1). A declaration made in a project that gives aliases to a name the global pool already has now registers the project's own entity, so other projects do not read those aliases. Aliases declared this way before 2.6.0b1 stay readable everywhere: the stored data does not record which scope declared an alias.
Smaller corrections (2.6.0b1): a store without associations no longer answers with an empty associations object; retract refuses true and false as ids; a quote carries a qualifier that starts on the next line; an episode's reconstruct quote is measured in the stored summary, without the [Episode] label; check_health reports and rebuilds node sets with a gap or a missing embedding.
session_key is at most 256 characters (2.6.0b2). Every tool that takes it declares maxLength: 256, and a longer key is refused before the call runs, with an input validation error. Keys minted from a process id and its start time, or a UUID, are well under the bound.
A listing the row cap cut says so (2.6.0b2). list_memories and list_episodes still clamp limit to 500 and 200 rows. When the caller asked for more and rows past the cap exist, the response carries budget_rows (the cap). A client that stops when it gets fewer rows than it asked for should read this key.
How far behind the vector index is counts every row a query reads from the table (2.6.0b2). python -m cpersona.vector_index status adds rows_read_exactly, excluded and unembedded, and build adds unembedded. check_health's vector_index_tail_grown compares rows_read_exactly against the rebuild ratio, so after fix=true fills missing embeddings it can appear with rows_past_watermark at 0. Rebuilding the index takes in the rows that have since gained an embedding.
Smaller corrections (2.6.0b2): on an in-memory database, export_memories reads a consistent copy; the vector index path scores a window it has to copy one chunk at a time, in the ranges the table scan uses.
A reconstruct budget below one head quote is raised to it (2.6.0). With the default CPERSONA_RECONSTRUCT_QUOTE_CHARS of 800, a smaller budget becomes 800 and budget_policy says so. It used to be raised only to 500, while the head could run to 800 without the overrun being reported.
A recall excerpt reads the block index only while block retrieval is on (2.6.0), as reconstruct's quote does. With CPERSONA_BLOCK_BUILD_ENABLED=true and CPERSONA_BLOCK_RETRIEVAL_ENABLED=false, excerpt_basis is now lexical.
A block set is current only if it respects the node layout and has every re-rank vector (2.6.0). Where block building is on, sets built before their record's nodes existed are rebuilt by the sweep and by check_health(fix=true) after the upgrade.
A non-default CPERSONA_PRIOR_FAR_WEIGHT is part of the calibration (2.6.0). A deployment that sets it to anything but 1 recalibrates on its first start of 2.6.0, because a gate measured at another weight is no longer restored. At the default nothing changes.
A time cue searches past the scan window through the coarse index by default (2.6.4a2). With CPERSONA_CUE_COARSE_ENABLED unset, the part of a cue's period past the scan window is searched when a coarse index exists, and otherwise left out with time_cue.remainder saying so; nothing reads every stored vector. A store within the window sees no change. true and false keep their meaning (settings).
Rows held beside the answer earn no recall count (2.6.0). With confidence enabled, the block reservation, the time cue's seats and the propagation seat no longer raise recall_count, and their confidence reads their own history.
Smaller corrections (2.6.0): a reconstruct quote cut to the preview tier hands over the whole range it began in expand; shortfall_reason blames the budget only when it cut an item of the window; bounds.reached counts only the rows inside recall's limit; a time_cue past the representable range is clipped rather than raising, and a unit that is not a string is refused; a failing node or block build no longer holds the other queued tasks; in api embedding mode the node check no longer reads every record's text.

New in 2.6 and inert unless asked for: the reconstruct tool, the recall trace (trace=true), the time cue (time_cue), associations declared with declare_associations or on store, and the propagation seat (CPERSONA_RECALL_PROPAGATION_SEAT).

Checking the result¶

check_health(agent_id="<id>", checks=["missing_nodes"]) reports no records left to build.
deep_check does not report stale_scoring_version.
A few recalls you know the answer to return it.

Going back to 2.5¶

Restore the backup you took before the upgrade (the database and the calibration file), then install 2.5.12. Do not point 2.5 at a migrated database. 2.5.12 opens a newer schema only with a warning that it may misread it; nothing converts a schema version 17 database back.