Optional Evidence-First Answering

answer_from_evidence is a TypeScript library adapter above memory retrieval. It works with caller-supplied evidence and a caller-owned language model. It does not change createMemory, ingestion, recall, HTTP routes, MCP tools, or model configuration. There is no built-in hosted provider, automatic search, retry, training step, or extra verifier model.

Use it for evidence-grounded questions about documents, projects, personal memories, historical records, or fictional material. It is not an unrestricted chat or creative-writing API: answers require cited evidence. Existing recall remains available for applications that need different answering behavior.

Use Authorized Recall Results

import {
    answer_from_evidence,
    type evidence_reader_model,
    type long_memory,
    type public_recall_query,
} from 'longmemory';

export async function answer_authorized_query(
    memory: long_memory,
    model: evidence_reader_model,
    scoped_query: public_recall_query,
    signal?: AbortSignal,
) {
    const recalled = await memory.recall(scoped_query);
    if (!('context' in recalled)) throw new Error('Recall did not return an evidence context');
    return answer_from_evidence(
        {
            question: scoped_query.text,
            sources: recalled.context.evidence,
            knowledge: 'evidence-only',
            signal,
        },
        model,
    );
}

Construct scoped_query using your application's authenticated identity, permissions, selected world, and temporal scope. Do not let an untrusted caller choose another user's scope. The adapter itself has no database access and does not authorize supplied sources. Only pass evidence the user may access and that your selected model is allowed to process; apply redaction and retention rules before calling it.

You can instead supply your own evidence objects. Each has a unique id and text, with optional speaker, observed_at (Unix milliseconds), and status (current, superseded, or contradicted). Other fields are stripped before generation, including arbitrary metadata. IDs and quotes are preserved exactly. Returned references refer to these supplied evidence objects; when the text is a multi-source bundle, keep your original bundle provenance for further inspection.

Model Contract

Provide an object with generate(request): Promise<{ text: string }>. The request contains system, user, json: true, max_tokens, and an AbortSignal. Map these to your model client's corresponding options and return its response text. Local and hosted models use the same interface; no SDK is required by the adapter. Use a model that can follow structured-output instructions.

The callback must disable tools/browsing, use the supplied prompts without injecting private ambient context, and honor the output limit and abort signal. The adapter makes at most one callback invocation, but cannot prevent retries or tool calls performed internally by your callback. Keep that transport policy explicit. The adapter will stop waiting on timeout/cancellation even if the callback ignores the signal; it cannot force that provider to stop computing or billing. Its post-response size check likewise cannot prevent upstream allocation or token spending, so enforce streaming/response limits in your client too.

Policies and Limits

Option	Default	Behavior
knowledge	evidence-only	No separate general-knowledge inference; deductions may use supplied evidence
knowledge: allow-general-inference	Opt-in	Allows recommendations/inference based on grounded premises, with an inference field separate from source-backed facts
output_language	Question language	Model instruction; source excerpts remain in their original language
question_time	Unspecified	Optional Unix-millisecond anchor for relative-time questions
max_context_tokens	2048; maximum 32768	Estimated allowance for serialized sources, not the whole prompt or a provider-exact token count
max_output_tokens	2048; maximum 8192	Requested limit passed to the model callback
max_response_chars	32768; maximum 131072	Rejects larger returned strings before JSON parsing
timeout_ms	60000; maximum 300000	Per-call timeout; combined with caller cancellation

Up to 64 sources are accepted, each at most 262144 UTF-16 code units. Sources are considered in supplied order. Oversized sources are omitted whole, not silently truncated; later smaller sources may still fit. used_source_ids means sent to the model, not necessarily cited. omitted_source_ids identifies skipped sources. No model call occurs if there is no usable evidence. Retrieval is the caller's responsibility, and the adapter does not fetch omitted or missing data.

The model returns a small ledger of subject/relation/object facts with exact source excerpts, a final answer, optional general inference, and missing-evidence descriptions. These are declarative facts, not a reasoning transcript. The output allows at most 32 facts with eight citations each; excessively large schemas fail closed. This is intentionally bounded, not an exhaustive whole-corpus report API.

Results and Errors

• answered: a nonempty answer and at least one fact whose citations pass the reference/excerpt checks. missing_evidence is empty.
• insufficient_evidence: answer and inference are null; missing evidence or ambiguity is described. Partial cited facts may be returned. Applications can render their own localized abstention; no fixed English user-facing answer is imposed. The no-source diagnostic message is internal English text.
• invalid_response: malformed, oversized, inconsistent, or unsupported output. answer is null and draft facts/inference are discarded. validation_errors contains a reason code such as invalid_json, invalid_schema, unknown_source, or quote_not_in_source. There is no automatic repair call.

Invalid requests and duplicate source IDs throw. Model errors, timeout, and caller cancellation also reject the promise; they are not reported as successful abstentions. Handle errors at your application boundary and do not expose provider error details or private evidence in public logs.

citations_verified means only that every returned citation refers to a source actually sent and that its quote is an exact contiguous excerpt. It does not verify semantic entailment, source reliability, completeness, actor attribution, arithmetic, language compliance, or the truth of the answer. A model can still misinterpret a valid quote or leave unsupported claims in its answer. Evidence is serialized as data and the prompt rejects source instructions, but this is not a prompt-injection-proof security sandbox. Treat all returned strings as untrusted when rendering them. Use appropriate review for consequential decisions.

The adapter never writes facts, infers supersession, changes memory contracts, or merges event identities. Those remain core/application decisions. Automatic follow-up retrieval is intentionally absent; a future bounded extension needs independent evaluation and must preserve the original authorization and time scope.