Problem
Company assistants hallucinate or leak context when retrieval, synthesis, and permissions are treated as one loose step.
Company memory that cites its sources, or says there is not enough verified information.
Project information
- Category: Company memory · Grounded Q&A · Access-aware retrieval
- Focus: Retrieve, deliberate, cite, or refuse unsupported claims
- For: Teams who cannot accept invented company facts
- Role: Reference system behind reliability engagements
Overview
Agents invent company facts when memory is weak. Agent Brain keeps answers tied to evidence, and can say “not enough verified information” instead of improvising.
The surface teams get is memory they can use for decisions, agent context, and operator Q&A, without treating every fluent answer as truth.
Work with me:
Mechanism
Retrieval, reasoning, citation, abstention, and access-aware filtering are separated into visible control points.
What It Proves
Memory can stay useful under scrutiny when “not enough evidence” is a designed outcome, not a failure.
Engagement Relevance
Useful when teams need grounded answers, operator confidence, and clear boundaries around what agents may know.
CTA
Use this pattern to assess or harden company-memory agent workflows.
Send a problem briefTechnical evidence
LongMemEval (latest full prod run): retrieval recall@5 on a memory corpus (k=5).
recall_at_k: 0.988 (98.8%)
recall_at_k_ex_dataset_abstention: 0.9893 (98.93%)
hits / scored @k=5:
hits: 493 / 499
hits_abstention_subset: 464 / 469Known limitation: This is retrieval evidence (finding supporting material). Cite-or-abstain decisions still depend on downstream citation/governance gates and access-aware retrieval.
Goal
Institutional memory that is useful and accountable. Grounded answers. Visible sources. Access-aware recall. Abstention when the corpus cannot support a claim.
What it does
- Grounded responses: claims point back to supporting material when they are made.
- Cite or abstain: refusal beats confident invention when evidence is thin.
- Access-aware memory: what a principal can see shapes what can enter an answer.
- Operator clarity: gaps and exclusions show up in the experience, not as silent failures.
Trust in practice
Memory products earn trust by what they refuse to say, not only by retrieval quality on easy questions.
- Evidence before prose: synthesis is a controlled step, not an open chat dump over documents.
- Source discipline: teams can see what was used versus held back for policy reasons.
- Evaluation mindset: reliability is evaluated and reported, not assumed.
Where it fits
- Use cases: company Q&A, agent context packs, meeting prep, policy-aware assistants.
- Signal: especially useful when hallucination or cross-team data leakage is unacceptable.
What I built
I built Agent Brain around a clear contract: cite, abstain, and respect access. Not an unbounded chatbot over a document dump.
Why it matters for clients
Institutional knowledge stays useful under scrutiny when the system stays honest when evidence is incomplete. That contract is what pilots and retainers put in place.