Agent Memory for Research Automation

Facts: Statements from papers
  "Paper 2314: CRISPR editing has 3% off-target rate in mouse models"
  "Paper 2198: Off-target rates vary by guide RNA design, up to 15%"

Events: Tests the agent ran
  "Hypothesis H1 tested: off-target rates scale with guide length"
    → Result: CONFIRMED (papers support 5–15% range)
  "Hypothesis H2 tested: off-target rate < 1% achievable"
    → Result: REJECTED (no papers show < 1% rate)

Entities: Concepts, researchers, methods
  "CRISPR-Cas9": A gene editing technique, standard in 2020+
  "Off-target effects": Unintended cuts at similar sequences
  "Gournay, M.": Researcher specializing in guide RNA design

1. Retrieve similar papers (semantic: "off-target rate")
   → Agent sees all prior work on this topic

2. Retrieve tested hypotheses (keyword: "off-target rate", causal: "REJECTED"?)
   → Agent checks: "Have I tested this before? Did it fail?"

3. Retrieve contradicting facts (semantic: topic, with logic: facts that disagree)
   → Agent surfaces conflicts: Paper A says 3%, Paper B says 15%

4. Retrieve related entities (graph: researchers, methods, concepts)
   → Agent finds experts and standard methods relevant to this work

import { Brain } from "brain-ai";

const memory = new Brain({
  namespace: "research",
});

// Research agent reads paper and stores findings
async function processPaper(paperId: string, content: string) {
  // Extract facts from paper
  const facts = await extractFacts(content);
  // e.g. "CRISPR off-target rate typically 3-10% in mice"

  for (const fact of facts) {
    // Store fact
    await memory.store({
      agentId: "researcher-1",
      content: fact,
      type: "fact",
      metadata: {
        paperId,
        topic: extractTopic(fact), // "off-target-rate", "crispr", etc.
        confidence: "high", // Based on study design
        timestamp: new Date().toISOString(),
      },
    });

    // Before trusting this fact, check for contradictions
    const contradictions = await memory.retrieve({
      agentId: "researcher-1",
      query: fact, // Semantic: similar statements
      topK: 5,
      filters: { type: "fact" },
    });

    // Flag conflicts for agent to investigate
    const conflicts = contradictions.filter(
      (c) => contradictsWith(fact, c.content)
    );
    if (conflicts.length > 0) {
      console.log(`CONFLICT: ${fact} contradicts:`, conflicts);
    }
  }

  // Before testing hypothesis, check if already tested
  const hypothesis = "Off-target rates scale with guide length";
  const priorTests = await memory.retrieve({
    agentId: "researcher-1",
    query: hypothesis,
    topK: 3,
    filters: { type: "event" },
  });

  if (priorTests.length > 0) {
    const outcome = priorTests[0].metadata?.outcome; // "CONFIRMED" or "REJECTED"
    if (outcome === "REJECTED") {
      console.log(
        "Hypothesis already tested and rejected. Skipping redundant work."
      );
      return;
    }
  }

  // Only test if not already tested or outcome was unclear
  const result = await testHypothesis(hypothesis);

  // Store test result
  await memory.store({
    agentId: "researcher-1",
    content: `Tested: ${hypothesis}. Result: ${result.outcome}. Evidence: ${result.evidence}.`,
    type: "event",
    metadata: {
      hypothesis,
      outcome: result.outcome, // "CONFIRMED" | "REJECTED" | "UNCLEAR"
      evidenceCount: result.evidence.length,
      timestamp: new Date().toISOString(),
    },
  });
}

Paper 1 (Summary stored):
  Fact: "Off-target rate ~3% in mouse models (Jinek et al., 2016)"

Paper 50 (Summary stored):
  Fact: "Off-target rate 15% with certain guide RNAs (Doudna et al., 2020)"

Agent memory retrieval when reading Paper 500:
  Query: "What do we know about off-target rates?"
  Results:
    - Paper 1: 3% (mouse, Jinek)
    - Paper 50: 15% (human cells, Doudna)
    - Paper 200: 8% (average across studies)

Agent notices CONFLICT: "Off-target rates vary from 3% to 15%"
  Queries memory: "What explains the variation?"
  Finds: "Guide RNA design explains variation (Paper 50)"
  Finds: "Cell type explains variation (Paper 87)"

Agent generates hypothesis: "Off-target rate = f(guide_design, cell_type)"
  Checks memory: "Have I tested this?"
  Memory says: NO
  Agent tests hypothesis against 1000 paper summaries
  Result: CONFIRMED (explains 80% of variance)

interface ResearchMemory {
  agentId: string; // "researcher-1"
  namespace: "research";

  // Content
  content: string;

  // What is this?
  type: "fact" | "event" | "entity";

  metadata: {
    // For facts: source, confidence, topic
    paperId?: string;
    topic?: string; // "off-target-rate", "crispr", etc.
    confidence?: "high" | "medium" | "low";

    // For events: hypothesis, outcome, evidence
    hypothesis?: string;
    outcome?: "CONFIRMED" | "REJECTED" | "UNCLEAR";
    evidenceCount?: number;

    // For entities: type, related topics
    entityType?: "researcher" | "method" | "concept";
    relatedTopics?: string[];

    timestamp: string;
  };
}