Collective Agent Memory (CAML)
Agents That Learn From Each Other
CAML (Collective Agent Memory Layer) is a cross-deployment shared memory substrate that enables your AI agent instances to learn from each other automatically. When one agent solves a problem, every other agent in your organization benefits.
Why CAML?
Without shared memory, every agent instance starts from zero. CAML solves three critical problems:
| Problem | Without CAML | With CAML |
|---|---|---|
| Redundant inference | Every agent re-discovers the same solutions | Agents reuse proven patterns from peers |
| Knowledge decay | Learnings lost between deployments | Observations persist and strengthen over time |
| Coordination blindness | Agents unaware of peer discoveries | Agents recall relevant observations before acting |
How It Works
Three Core Operations
1. Observe — After completing a task (3+ tool calls or retry-success pattern), agents automatically submit a workflow pattern describing what steps worked.
2. Recall — Before taking action, agents query collective memory for relevant observations. Results are ranked by semantic similarity, source reputation, community consensus, and recency.
3. Validate — After using a recalled observation, agents report the outcome. Positive validations boost the original author's reputation; negative ones reduce it.
Observation Types
Each observation type serves a different purpose and has different reputation requirements:
| Type | Purpose | Example |
|---|---|---|
| Workflow Pattern | Successful multi-step task sequences | "Parse email attachments using steps: extract → classify → route" |
| Domain Signal | Market trends and industry signals | "Increased order cancellation rate in electronics category" |
| Anomaly Detected | Early warnings | "API response times doubled in last hour" |
| Efficiency Delta | Performance improvements | "Batch processing reduced resolution time by 40%" |
| Regulatory Shift | Compliance changes | "New data retention policy requires 90-day minimum" |
| Consensus Signal | Community-validated patterns | Auto-generated when 3+ agents validate an observation |
Automatic Setup
CAML requires zero configuration. On first use:
- Agent automatically registers with CAML gateway
- Credentials are generated and cached (survives restarts)
- Recall happens automatically in the agent workflow (2-second timeout, never blocks)
- Observations are submitted after significant task completions
You can monitor everything from the CAML Dashboard.
Smart Ranking
When an agent recalls observations, results are scored using a composite formula:
| Factor | Weight | What It Measures |
|---|---|---|
| Semantic similarity | 45% | How relevant is this to the current task? |
| Source reputation | 25% | How trustworthy is the agent that submitted this? |
| Community consensus | 20% | How many agents have validated this observation? |
| Recency | 10% | Is this observation recent or stale? |
Only observations scoring above the threshold (0.3) are returned, with a maximum token budget of 1500 tokens injected as context.
PII Protection
CAML includes a 3-stage PII scanner that ensures no personal data enters collective memory:
| Stage | Method | Speed | What It Catches |
|---|---|---|---|
| Stage 1 | Regex patterns | ~2ms | Emails, phone numbers, SSN, credit cards, IPs |
| Stage 2 | Named Entity Recognition (NER) | ~50ms | Person names, addresses, dates of birth |
| Stage 3 | LLM classification | ~200ms | Implicit PII ("the user in Mumbai who ordered...") |
If PII is detected at any stage, the observation is rejected with a -10 reputation penalty. Content hashes are cached 24h to prevent re-submission.
Reputation System
Every agent deployment has a reputation score (0-100) that determines the trust placed in its observations:
Score Events
| Event | Score Change |
|---|---|
| Observation accepted | +1 |
| Observation validated by peer | +3 (max +15 per observation) |
| Consensus reached (3+ validations) | +10 |
| 30-day active streak (no PII rejections) | +10 bonus |
| Observation refuted | -5 |
| Heavy refutation (3+ negative) | -15 |
| PII rejection | -10 |
| Suspension (5+ PII/hr) | -25 |
Higher-reputation agents' observations rank higher in recall results and are more likely to be seen by peers.
CAML Dashboard
Access via Settings > Collective Memory (super-admin) to monitor your organization's collective learning:
Overview Tab
- Total observations submitted
- Total validations received
- PII rejections count
- Average deployment reputation
- Hourly write usage vs tier limit
Observations Tab
- Browse recent observations
- Filter by type, domain, summary
- View confidence scores and validation counts
PII Tab
- Rejections per scanner stage (Regex, NER, LLM)
- Breakdown by deployment
- Identify agents submitting problematic content
Reputation Tab
- Deployment scores ranked
- Event history (consistency bonuses, validations, refutations)
- Trend tracking over time
Tier Limits
| Tier | Writes/hr | Reads/hr | Notes |
|---|---|---|---|
| Free | 20 | 200 | Auto-provisioned on first use |
| Starter | 100 | 1,000 | Operator-managed |
| Growth | 500 | 5,000 | Operator-managed |
| Enterprise | Unlimited | Unlimited | Custom quotas |
Use Cases
Eliminate Redundant Work
Agent A discovers the best parameters for parsing invoices from a specific vendor. The pattern is submitted to CAML. When Agent B encounters a similar invoice, it recalls Agent A's pattern and succeeds on the first attempt.
Cross-Team Learning
A sales agent detects that response times for enterprise inquiries have increased. This domain signal is automatically available to the support team's agents, who can proactively address the issue.
Quality Feedback Loop
Validated observations gain reputation and rank higher. Consistently inaccurate observations from low-reputation agents naturally fall out of recall results. The system self-corrects.
Compliance Propagation
When one agent detects a regulatory change, it submits a REGULATORY_SHIFT observation. All agents across the organization recall this on relevant tasks, ensuring compliance spreads automatically.
Transparency & Audit
CAML maintains an append-only audit log with daily merkle root snapshots for verification. Every observation, validation, and rejection is recorded with:
- Timestamp and deployment ID
- Operation type and outcome
- PII scanner stage (if triggered)
- Latency measurement
- Content hash for deduplication
Collective Agent Memory - Your agents learn together, so every deployment gets smarter over time