Live memory bindings#

Bind a fresh conversation slot to a typed GPU store when its prompts must use scoped selected memory. A binding fixes lineage, principal, room, conversation identity, scope and page cap. It cannot replace an existing binding or adopt a slot with earlier turns. Unbound conversations keep their transcript-based input path.

Choose a retention mode#

Mode 2 requires an exact accepted model, wrapper and processor combination. See semantic memory qualification. Text query preparation also requires the embedding model; prepared queries supply their own vectors.

Commands and prerequisites#

The local operator can use these commands from standard input or an attached input file:

memory load PATH
memory apply PATH
memory bind PATH
memory query PATH
memory text PATH
memory retain PATH
memory resume PATH

resume restores a live memory checkpoint, including its required conversation bindings.

load reads the verified checkpoint and tail from a CCIR file. apply reads a canonical typed tail. bind and query read raw batch files with the layouts below. A path is the rest of the line, with spaces kept as path bytes. There is no shell expansion or quote processing.

text reads the same batch shape with zeroed vector fields and a 2,048-byte input limit. The GPU prepares those vectors with the loaded embedding model. Text requests defines its required model roles and exact byte layout.

retain names the last accepted input of each idle binding. The device retains its exact text and prepared vector. See Retain accepted input for the request, recorded mutation and optional working focus.

Files must be bounded regular files. The reader refuses symbolic links, pipes, directories, incomplete reads and extra bytes after the declared file size.

Read and framing refusals appear on standard error and as a console note.

The disk reader checks framing and transmits exact bytes. The device checks the full batch before it changes state. It checks IDs, versions, bindings, ordinals, scope, the declared store cut and the prepared embedding space. The reader does not search memory or construct prompts. memory choice is not an input command for typed choices.

Byte layout#

All integers are unsigned and little-endian. IDs are 16 bytes. The shared constants and offsets are in cuda/cognitive/live.h. A batch has 1 to 64 rows. One query batch waits for its recorded choice before the next typed transfer is admitted.

Each class A record of type 33 has a 32-byte prefix and at most 160 data bytes:

Parts are in order. Each part has 160 data bytes except the final part. The feeder publishes at most 32 records per group and waits for ring space between groups. Each file transfer gets a new random ID. A load has two 64-bit lengths at offsets 0 and 8, then the exact checkpoint bytes and exact tail bytes.

An update is one typed tail. The maximum load is 16 + 2 * image_bytes, using the compiled image capacity. It is 37,749,008 bytes with the default capacity, including the 16-byte length prefix.

Bind, query and choice have a 64-byte header. Their magic bytes are AOTXBND1, AOTXLIV1 and AOTXCHO1. Count and schema are 32-bit fields at 8 and 12; lineage is at 16. The 64-bit store sequence is at 32, and the 32-bit row size is at 40. Bind and query bytes 44 through 63 are zero. A choice has a 32-bit status at 44 and zero bytes 48 through 63.

Status 0 means success; a nonzero status from the typed status table is a refusal with count 0 and no rows. A choice has the same transfer ID as its query. Only the device emits it.

Slots, scope, page caps and automatic retention are 32-bit values. Automatic retention is 0 for explicit retention, 1 for source retention or 2 for qualified semantic intake. See automatic input retention for combined decisions and admission. The prepared query and selection layout are defined in prepared memory.

Ordinals start at 1 for each binding. The largest bind is 4,160 bytes; the largest query is 528,448 bytes; the largest choice is 37,952 bytes. A choice keeps zero bytes in unused selection slots.

The 32-bit focus flag is 0 for caller references only. Flag 1 appends the binding's working references after explicit focus, as defined in retention.

State, prompts and recovery#

The resident typed store uses the configured capacity. The defaults are 8,192 immutable object versions and 16 MiB of payload per store. Explicit updates require idle cognitive conversations. Each binding keeps bounded current request and choice state. Later queries replace this current state; the journal keeps prior requests. The 64-pair saved-query limit of the offline recall file does not limit the number of live turns.

Prepared queries supply their own vectors; text requests use the GPU embedding service. Explicit retention creates memory from accepted input. Automatic bindings retain each input during admission. Other memory changes require typed operations. Cold memory supplies an explicit offload and retrieval path; binding alone does not enable it.

The device uses the model's loaded prompt format, labelled selected memory and current input. It checks context and page limits. Tool continuations retain the bound context with the current call and result, and check selected dependencies again. Bound prompts do not take their memory from the audit transcript.

Load, update, bind, query and choice bytes are in the class A journal. Choice output is limited to 64 records per tick and completes before prompt admission. Restore uses the recorded choices without a new search. It does not need the source CCIR or request paths. Incomplete transfers cannot publish state or prompts. Missing or invalid choices at restore completion refuse affected work.

Each choice completes one pending query. Repeated choices refuse restore.

The journal preserves this interface's authoritative requests and choices. Use live checkpoints for a continuous memory mirror, or complete runtime files for packaged assets and history.

Per-agent audit JSONL contains the exact accepted input and a selection row. The selection names lineage, conversation, principal, room, scope, full ordinal, request ID, selection ID, store cut and selected object IDs with versions. A complete matching device choice is required before accepted input is written. A refused choice produces a selection status row for a safely framed request, without an accepted input row. Malformed or incomplete batches do not become accepted text. These files are derived audit data.

Typed audit turns follow the last device manifest, including prior tool turns. Ordinary input does not advance typed audit turns. The request ordinal is separate from the model turn number.

See memory maintenance for retention policies, GPU reclamation and automatic file shrinking.

Extended operations#

The prefix table above covers the original live request family. Operation 12 records direct admission, 13 carries maintenance and 14 carries semantic results. Appraisal, cold memory and task reviews define their additional operation layouts. Unknown operations or required schemas are refused.