[Integration]: Governance-aware agent memory with Hindsight

[nagasatish007]

Summary

Hindsight provides agent memory that works like human memory — with importance-based retention, forgetting curves, and contextual recall. TealTiger governance decisions are a natural fit for Hindsight's memory system: governance events have varying importance, should decay over time, and need to be recallable during future decisions.

Why this integration?

Currently, TealTiger governance decisions are either in-memory (lost on restart) or persisted in flat storage. Hindsight adds intelligence to governance memory:

• Importance-based retention — DENY decisions (high risk) persist longer than routine ALLOWs
• Contextual recall — "What governance decisions were made for this agent in similar contexts?"
• Forgetting curves — Old, low-importance approvals naturally age out
• Pattern detection — Recall clusters of denials to detect systematic policy violations

Proposed approach

1. HindsightGovernanceMemory — stores TealTiger decisions in Hindsight with importance scoring
2. Decision importance mapping: DENY=0.95, ALLOW=0.60, MONITOR=0.70 (mirrors Dakera pattern)
3. Contextual governance recall — before making a new decision, recall similar past decisions
4. Anomaly detection — flag when an agent's denial rate spikes vs. historical baseline

Example usage

from hindsight import HindsightClient
from tealtiger.integrations.hindsight import HindsightGovernanceMemory

hindsight = HindsightClient(api_key="your-key")

governance_memory = HindsightGovernanceMemory(
    client=hindsight,
    importance_map={
        "deny": 0.95,
        "allow": 0.60,
        "monitor": 0.70,
    },
)

# Every governance decision is stored with importance-weighted retention
client = TealOpenAI(
    api_key=os.getenv("OPENAI_API_KEY"),
    guardrails={"pii_detection": True},
    on_decision=governance_memory.store,
)

# Later: recall governance history for context
history = governance_memory.recall(
    agent_id="research-agent",
    context="tool call to web_search",
    limit=5,
)

Deliverables

• tealtiger.integrations.hindsight module
• HindsightGovernanceMemory class implementing importance-weighted storage
• Contextual recall for governance-aware decision making
• Example: agent that learns from past governance denials
• PR to Hindsight cookbook showing TealTiger integration

Links

• Hindsight repo: https://github.com/vectorize-io/hindsight
• Hindsight cookbook: https://github.com/vectorize-io/hindsight-cookbook
• TealTiger Dakera integration (similar pattern): https://docs.tealtiger.ai/integrations/dakera/

[nagasatish007]

vectorize-io/hindsight#2284
vectorize-io/hindsight#2770