MemoryLake 用例集：12 个基于记忆层的 AI 应用场景演示

sales-call-memory Sales & revenue teams Call transcripts → extracted, dated facts; rep notes pinned as facts; hand-off docs as documents; a CSM brief generated by searching the account memory. CLI runner plus a local web UI (python3 web/server.py) ~3 min CLI 4:12 · Web 4:59 chatbot-user-memory Engineering & developers Per-user memory for a chatbot: sessions as user↔bot conversations, facts owned by the user's actor, one scoped search per turn as the model's memory block, cross-user isolation check, right-to-be-forgotten via fact delete. CLI runner plus a local web UI ~2 min CLI 2:40 · Web 2:49 research-memory Research & analytics A quarter of research reused next quarter: PDFs and an Excel model imported as documents (search returns the file type and the sheet), model reviews as conversations, a revised assumption kept as one fact with both values and both dates, the original file downloaded back from memory. CLI runner plus a local web UI ~3 min CLI 3:51 · Web 4:03 meeting-memory Operations, HR & teams One meeting memory across five tools: Otter, Granola and Slack meetings as group conversations (one actor per speaker), Notion and Fathom exports as Word documents, a decision that changed kept as one fact with both dates, action items pinned to their owners and found by actor tag, a confidential call deleted and gone from search. CLI runner plus a local web UI ~4 min CLI 4:50 · Web 5:35 codebase-memory Engineering & developers One memory per repository: each repo's docs/ tree (ADRs, runbook, a PowerPoint review deck) mirrored into Library folders and imported with one --recursive command, PR and incident reviews as conversations, a pre-commit check that blocks a pattern only in the repo that banned it, and day-one questions answered across both repos with every hit labelled by repo. CLI runner plus a local web UI ~6 min CLI 6:36 · Web 7:35 contract-review-memory Legal & IP Contract review that remembers your positions: standard positions pinned to a playbook project, each counterparty's accepted exceptions pinned to its own actor, one search scoped --projects <playbook> --actors <counterparty> that brings back ours + theirs and nobody else's, review sessions with the redline attached as a FILE block, last year's session found by metadata and replayed, its redline downloaded back and diffed against the renewal. CLI runner plus a local web UI ~5 min CLI 5:41 · Web 6:48 onboarding-memory Operations, HR & teams Onboarding memory that answers with the policy in force: every version of every policy pinned as a fact stamped with its effective date, sign-off and source; a new hire's search returns old and new versions and the stamps pick the one in force (stale versions and the outdated handbook marked); an audit of what applied on any date via fact list; new hires moved to stage:day-1 with actor update, found with actor list --tags, each getting the first-day plan for the role in their metadata. CLI runner plus a local web UI ~3 min CLI 3:17 · Web 4:35 customer-onboarding-memory Customer success & support One memory per customer across the onboarding handoffs: sales call, kickoff and an implementation chat as conversations tagged by stage, an AI assistant's tool calls stored as TOOL_USE / TOOL_RESULT blocks, a trainer's discovery checklist answered from memory (8 of 8, against 4 in the handoff notes), the tool findings nobody passed on replayed by tool_call_id and promoted into memory (0 → 1 hits), and a timeline built from per-message metadata. CLI runner plus a local web UI ~4 min CLI 4:51 · Web 6:02 candidate-pipeline-memory Recruiting & HR One memory per candidate across every stage: screen, technical and values interviews as candidate↔interviewer conversations, each interviewer's scorecard pinned to the candidate, a debrief that labels every answer by stage (a later interview revises an earlier answer in place), a new recruiter searching the whole pipeline from the workspace's ACTIVE bindings, a withdrawal that actor unbind seals (reads, searches and writes refused, actor kept) and a re-application that actor bind restores with identical fact ids. CLI runner plus a local web UI ~5 min CLI 5:47 · Web 7:10 agent-bug-replay Engineering & developers Reproduce an AI agent's bug by replaying its memory: four support chats, the agent's memory_search call and the facts it got back stored as TOOL_USE / TOOL_RESULT, the same call re-run today (different answer), then memory pinned to the exact message with fact trace (ADD / UPDATE / FORGET, dated by source message) and conv fact-actions — 6/6 facts identical to what the agent recorded — plus a then-vs-now diff with a source on every line and an audit record. CLI runner plus a local web UI ~5 min CLI 6:24 · Web 6:25 brand-asset-memory Design & creative Brand assets the AI actually sees: six DAM images with meaningless names imported as image_file documents, each described by MemoryLake with the text read off it and the questions it answers; a designer's plain-words questions answered with the right picture at rank 1 (judged ✓/✗ against the file the brand lead would send) and the original downloaded into a brief; a sign-off chat with the pictures as IMAGE blocks; a retired wordmark removed and no longer returned. CLI runner plus a local web UI ~3 min CLI 4:06 · Web 5:04 preference-memory Operations, HR & teams The AI that actually knows you, and keeps only what you allow: sessions from two tools land on one person's memory, the built-in default keeps her family, diet and surgery, she sets her own fact instruction (draft → set), and the next session's conv fact-actions receipt shows the new work rules added, the changed answer length updated in place, and 0 facts in each excluded category (✓/✗ per category, measured 37 of 42 sessions clean); the profile then goes byte for byte into Claude, OpenAI and Gemini request payloads. CLI runner plus a local web UI ~2 min CLI 3:45 · Web 4:02 conflict-check-memory Legal & IP Conflict check memory for a small law firm: the firm history as a project (matter index pinned, archived letters imported) plus one tagged actor per party; each intake gets a party check (per-party memory + firm-history search, judged by the firm's rules) and MemoryLake's contradiction detector (fact conflict), which flags the intake form's “no prior relationship” claim against the 2019 matter record (m2m) and the closing letter, quoting it (m2d) — 8 of 8 runs, 0 false positives in 24 checks; resolved with trust_document / keep_fact, and fact conflict get becomes the audit trail. CLI runner plus a local web UI ~5 min CLI 6:25 · Web 8:06 verifiable-source-memory Engineering & developers Agents that cite verifiable sources: four PDFs (two versions of a fee schedule, a policy with a hold table, wire procedures) imported with their provenance and sha256 as Library attributes; every claim in the agent's (scripted) draft cited to the passage that states it — document, version, page and box on the page (search hits' highlight.chunks[].range), the sentence or table row verbatim — or to a member fact with its origin from fact trace (stated by staff vs extracted from her message); a claim only a superseded version supports is replaced, one nothing supports is dropped; every citation re-checked against the downloaded original (sha256 + that page's text layer, 9 of 9 in 7 of 7 runs, controls ✗) and exported as JSON. CLI runner plus a local web UI ~2 min CLI 3:10 · Web 4:53 macro-memory Customer success & support Support macros that stay true to the policy: 7 saved replies exported from Zendesk and Intercom, one fact each with its name, tool and the policy sections it depends on in metadata, next to the support policy as a document; MemoryLake's contradiction detector flags a stale macro against the policy, quoting it (m2d), a macro that contradicts itself (self) and two tools' macros that disagree (m2m); each settled with a different strategy — edit_fact, trust_document + version 2, trust_fact (the server rewrites the macro, so the approved wording is restored and fact trace shows all three versions) and dismiss on the conflict the fix made stale; then a new policy flags nothing until the dependent macros are re-filed, and the shipping macro is flagged against v4. CLI runner plus a local web UI ~6 min CLI 5:57 · Web 7:53 pitch-deck-memory Marketing & agencies A new pitch deck built from the slides that already worked: three past decks (.pptx) and their debriefs (.docx) in one project; for every slot (case study, credentials, process, pricing, team) search returns the decks one passage per slide (range = P<slide>[…]), and the slide is cited as deck + slide number + its own words, picked by category or by won pitches (Library attributes); debrief sentences, cited by page, reorder the outline; approved figures pinned with fact add get the old slide flagged by the contradiction detector (m2d, matched back to slide 4 of the .pptx); every citation re-checked against the downloaded .pptx/.docx (sha256 + that slide's XML, 8 of 8 in 7 of 7 fresh runs, neighbouring-slide control rejected) and exported with the outline. CLI runner plus a local web UI ~3 min CLI 3:55 · Web 4:52 buyer-profile-memory Sales & real estate One buyer, three agents: the buyer's stated profile (dealbreakers, budget, financing) pinned on her own actor, the showings as events on the team project; a covering agent's notes contradict the kickoff, and MemoryLake's contradiction detector flags both on the buyer's actor (fact conflict list --actor, m2m, 8 of 8 fresh runs); the new agent's tour screen marks the listings that depend on the dispute as ask first, the showings show what she said vs what she responded to, keep_fact settles each conflict with the showings as evidence, and the screen flips; fact conflict get keeps the forgotten notes word for word in the handoff record. CLI runner plus a local web UI ~3½ min CLI 4:08 · Web 5:14 reflexion-memory Engineering & developers A self-improving (Reflexion-style) agent whose reflections live in its own agent memory: five failed CI runs as conversations (incidents land on the project), a typed reflection after each (fact add --agent + metadata: cause, adjustment, expected outcome); MemoryLake flags the two lessons that contradict each other on the agent (fact conflict list --agent, 7 of 7 fresh runs) and keep_fact settles it; today's planning step retrieves the right lesson (search --actors <agent's actor>) while a twin made with agent fork (same config, no memory) repeats the mistake; two lessons with the same cause become a published skill (security review safe) on agent version 2, memory unchanged; fact trace + fact conflict get + version lists as the audit trail. CLI runner plus a local web UI ~4 min CLI 5:05 · Web 6:52 bookkeeping-memory Finance & accounting Bookkeeping memory per client: two clients, each a project with the owner's onboarding thread and past ledger exports (.xlsx); rules pinned from the threads (vendor, line, category, source message), so the same Paylane fee line is Cost of revenue for one client and Bank and merchant fees for the other; September's bank feed categorized by rule, else the last booking of that payee cited to its ledger row (file · Sheet!A15:E15, from the table passages search returns), else ask the client; proj doc inspect shows the tables found on each sheet (one sheet, two tables, title row recognized); a corrupted export (proj stats error 1, ERROR_FILE_CORRUPTED) fixed in place with lib upload --on-conflict overwrite + proj doc reload; every cited row re-checked against the downloaded workbook (next-row control rejected); and “why is it booked that way?” answered with the message and the batch MemoryLake read it in (conv consumed-messages). CLI runner plus a local web UI ~5 min CLI 6:17 · Web 7:48 tot-branch-memory Engineering & developers Branched memory for a Tree-of-Thoughts agent: three branches explored at once, each its own project with its own conversation; the reasoning in THINKING blocks, replayed verbatim but never turned into memory (0 of 70 reasoning-only details across 6 fresh runs; the same text as TEXT is extracted); three explorers race for one shared trunk log and the losers get 409 "not the current head" and retry on current_message_id, leaving a gap-free log; checkout = search --projects <trunk>,<branch>, merge = promote the winner's result and each pruned branch's reason to the trunk, rollback = proj delete (fact list → 404); a week later the next task reads the trunk first and opens 1 branch instead of 3, each skip citing the fact that says why. CLI runner plus a local web UI ~3 min CLI 3:59 · Web 4:41 security-review-memory Legal & IP An enterprise security questionnaire about agent memory answered with live checks against your own team: one key per integration (api-key create --name; the secret is shown once — an idempotent retry returns the same id without it, list / get show prefixes); api-key rotate (same id, old secret 401 on the next call, the new one reads the same facts); --expires-at (401 "has expired" right after, while status stays enabled); api-key revoke (a key cannot revoke itself: 409; revoked → 401 + get 404, what it wrote stays); an integration key is team-wide (caller_role: tenant_owner, reads another customer's memory — answered "no", with the mitigation); fact trace audits what / when / from which message but not which key ("partly"); offboarding proj delete → fact list 404, search 400. Writes a 7-answer evidence pack (5 yes · 1 partly · 1 no, 6/6 runs). CLI runner plus a local web UI ~3 min CLI 3:39 · Web 4:58 engagement-insight-memory Consulting & Strategy Lessons that outlive the engagement without the client in them: the same close-out debrief is filed in the engagement's project (built-in default: keeps the client, its people and its numbers) and in a firm-wide insight library with a project fact instruction (fact instruction set --project: describe clients by sector and size, never by name, no figures) — same words, two memories; a leak scan (every engagement's identifiers + a figure rule) checks each library fact (135 facts written under the instruction while building it, 0 named a client); a client figure pinned with fact add is stored word for word (an instruction steers extraction, not writes), ranks 1–3 for the next team's question, and is caught and deleted; the next engagement searches only the library; every lesson traced via fact trace → debrief messages → conversation metadata to an engagement code. CLI runner plus a local web UI ~3 min CLI 3:01 · Web 4:12 escalation-memory Customer Support Escalation know-how that stays when the senior agent leaves: a senior support agent talks her escalation notes (full of customers, contacts, ticket ids and a contract value) through with an escalation agent that has an agent-scoped fact instruction (fact instruction set --agent: record escalation paths, context packages and framing even when a teammate says them; never a customer, id or figure); the same words to a second agent on the built-in default keep almost nothing of what she said, and the project keeps the customers; leak scan over every fact (52 facts on the agent while building it, 0 named a customer); provenance via fact trace --agent → messages → speaker; her actor sealed with actor unbind (404/400) while the agent's lessons stay; the junior's three questions answered from the agent's memory via search --actors <agent's actor> (3/3 vs 0/3); his outcome adds a lesson, not the customer. CLI runner plus a local web UI ~3 min CLI 4:50 · Web 7:00 mcp-scope-memory Engineering One memory scope per AI tool: three tools (two editor assistants over MCP, a CI bot over Memory Router) each get a boundary — the memory store identifier: projects + one person + one agent (boundary create / get / list / update / delete); the same question through each returns three answers with every hit labelled by scope (0 from outside a boundary); a boundary for someone not on the workspace is refused at write time; after a project is deleted, a person sealed and an agent deleted, boundary get still lists all three — the bot's searches fail with HTTP 400 and the other two boundaries refuse even a rename — until an audit checks every reference against what exists and boundary update repairs each in place (same id). CLI runner plus a local web UI ~2 min CLI 3:41 · Web 5:14 patient-context-memory Healthcare Last week's allergy in front of today's order: a patient's chart on his own actor, every entry signed and sourced (metadata), read back page by page (--page-size / --continuation-token; a full last page still returns a token); urgent care records a cephalosporin allergy and the contradiction detector flags it against the 2018 "no known drug allergies" line (dated history is never flagged: it can all be true); keep_fact forgets every other fact in a conflict, so it retires the stale line on the two-fact conflict and the three-fact one is dismissed; the AI scribe's draft is checked with fact list --query (exact, case-insensitive, every page) for every name in the drug class — cefuroxime ✗ citing last week's entries, doxycycline ✓, and ✓ for cefuroxime against the chart as of the day before urgent care; the triage note says "Keflex", so the class name alone misses it, and search (semantic, top-k) is shown side by side; each decision is recorded with the ids and versions of what it saw, then re-verified. CLI runner plus a local web UI ~3½ min CLI 4:30 · Web 5:52 parent-communication-memory Education Every teacher's thread with a family, tagged as it is captured: six emails, calls, hallway chats, portal messages and a counselor note go into one Library folder per family with lib upload --xattrs (channel, who logged it, date, sensitivity, open follow-up); the default lib list returns none of the tags (only a pre-signed thumbnail URL), lib list --xattr-keys returns exactly the ones asked for; sensitivity routes each file into the teachers' project or a counselor-only one, with a leak check that removes a misfiled note; a one-call completeness audit before the parent meeting finds the teacher with no logged contact and 3 open follow-ups; the same question from a teacher and from the counselor (the separation note is #1 for her, absent for him); follow-ups closed after the meeting with lib xattr delete + resolved_on/outcome. CLI runner plus a local web UI ~2 min CLI 2:54 · Web 4:15 newsroom-memory Media Scanned records you can quote, and claims checked against them: a records request returns three image-only PDFs (0 fonts, no text layer); after proj doc import, proj doc inspect points at MemoryLake's OCR layout, read back line by line (30/30 lines word for word); a question returns the record, the page and its box, and the line (proj doc list --name finds a record by name); four interview claims pinned with who said them (fact add + fact update --metadata); the contradiction detector flags the spokesperson against page 1 of the inspection and the capital-plan memo (m2d) and against an unnamed engineer (m2m), with attribution on both sides (the excerpt's page found through the OCR layout); every quote in the draft checked against the page it cites, and a planted misquote caught. CLI runner plus a local web UI ~2½ min recording in progress