Text query preparation#

memory text PATH sends a batch of text requests to the live cognitive store. Load both model roles with --roles language,embedding. The device prepares query vectors with the loaded embedding model, selects memory and sends the context to the language model. The disk reader transports bytes and writes audit files.

Submit text#

Use an existing live binding and a correctly framed request file. memory text PATH is a console command, not a shell program. Each input has a limit of 2,048 UTF-8 bytes. Longer input is refused without truncation.

Prepared memory query requests have the same input limit. The source travels in ordered 160-byte fragments. A fragment can split a UTF-8 character; validation uses the complete source. A text request still names its binding, IDs, ordinal, store cut, scope, memory budget and optional required or focus references.

Input bytes#

Use the query envelope in live memory, with magic AOTXTXT1. The header is 64 bytes and each row is 8,256 bytes. A batch has 1 to 64 rows. In the 8,192-byte query part, these fields must be zero:

All other query fields retain their existing meaning. The device supplies the model digest, fixed processor digest, width and vector. The processor identity describes the current tokenization, pooling and normalization path. It does not state the quality of a model's semantic representation.

For sources of up to 192 bytes, the processor digest is SHA-256 of this exact ASCII line, without a line ending:

AOTX text embedding 1; exact UTF-8 1..192 bytes; model GGUF vocabulary; clean/pretok/merge/gather; all tokens from position zero; final row RMS output norm F32; L2 F32; cosine query F32; no instruction prefix

The digest in hexadecimal is:

7d12af1d2cd1e5194def983d1fd8073d1c36c444eea39c2dcf9bbe394e75892d

For longer sources, the exact identity line uses embedding 2 and 1..2048 bytes. Its digest is 3f5f1fdc067157f5b81e8818885bbba30484c233befe9d41aa77c037ceba5a36. The two identities use the same tokenization and numerical path. Recall accepts this named pair as compatible. Other processor identities still require an exact match. The ordinary tool and transcript text limits remain unchanged.

The feeder uses class A type 33, operation 6. Its regular-file checks, random transfer ID, exact byte transport and 160-byte fragments match the prepared-query path. Both standard input and attached input accept the command. Disk framing checks do not encode text or interpret the cognitive fields.

Recorded decision#

For bindings with explicit retention, only the device writes operation 7, with magic AOTXTCH1. Its header uses the same status and count fields as a prepared choice. Each successful row has 8,784 bytes:

The maximum decision is 562,240 bytes. A refusal has a 64-byte header, count zero and a nonzero typed status. It does not advance the ordinal or change the previous bound context. A text decision must match an outstanding text request and its transfer ID. A prepared choice cannot complete a text request, or the reverse.

The device records at most 64 fragments per tick. Prompt admission waits for the complete recorded decision. Embedding uses bounded temporary pages and a deadline of 128 service ticks plus the maximum batch token work divided into 512-token passes. For the 64-row profile, the deadline is 384 ticks. Each source retains its cache positions between passes. The service publishes only the vector from the final complete-source row.

A source waits for its complete cache extent before its first pass. Page contention uses the full batch deadline. The ordinary tool retry count does not shorten this wait.

Missing or incompatible model state, invalid input and resource failure refuse the full batch without a partial prompt. New text requests also refuse while a model load is pending.

An unserved page queue can delay cleanup. Prompt publication waits for queued page releases.

Recovery and audit#

Replay uses the recorded prepared query and selection. It does not tokenize, embed or search again. The embedding weight identity and processor identity must remain valid. The four device-supplied field ranges can differ from the original text request.

With live prefix flag 1, the device can also append working references after the exact explicit focus prefix. See Retain accepted input for that rule. Wrong request types, changed input, malformed vectors and incomplete decisions refuse recovery of the affected work.

The audit pairs each original text batch with its complete matching decision. It writes the original input, IDs, scope, ordinal and selected object versions to the existing per-agent transcript. Bad framing, changed input or a partial decision cannot create accepted input. Recorded vectors remain journal bytes; the audit does not construct a CPU memory index.

A binding with automatic retention records operation 10 and retains input during admission. Other bindings use memory retain PATH to retain the last accepted input and prepared vector. Binding value 2 also enables semantic intake. Other memory changes require explicit typed state operations. Base conversations keep their existing input path.