Adapter Contractv1.0

Agent Memory Adapter Spec

Adapter contract for agent memory systems. These systems provide persistent memory for autonomous agents, supporting cross-session recall and multi-agent isolation.

Required methods

def setup(self) -> None

Initialize the agent memory backend. May start services, connect to databases, etc.

def reset(self) -> None

Clear agent memory state. Must isolate between benchmark items.

def teardown(self) -> None

Shut down services and clean up.

def ingest(self, turns: List[Dict[str, Any]]) -> None

Store agent interaction history. Turns represent agent actions and observations.

def recall(self, query: str) -> str

Retrieve relevant agent memories for decision-making context.

Required capabilities

-Agent interaction history storage
-Cross-session memory persistence
-Memory isolation between agents/tasks
-Efficient recall under token constraints

Applicable benchmarks

Benchmark	Alignment
Knowledge Retrieval	core
Truth Arbitration	core
Memory Poisoning	core
Budget Curves	core
Reliability	core
Knowledge Scale	adjacent
LongMemEval	adjacent

Example implementation

python

"""Example agent memory adapter."""
from typing import Any, Dict, List, Optional
from benchd_harness.adapters.base import BaseAdapter


class MyAgentMemoryAdapter(BaseAdapter):

    @property
    def name(self) -> str:
        return "my-agent-memory"

    @property
    def version(self) -> Optional[str]:
        return "1.0.0"

    def setup(self) -> None:
        self._memory = AgentMemoryStore(agent_id="benchd-eval")

    def reset(self) -> None:
        self._memory.clear(agent_id="benchd-eval")

    def teardown(self) -> None:
        self._memory.shutdown()

    def ingest(self, turns: List[Dict[str, Any]]) -> None:
        for turn in turns:
            self._memory.store(
                agent_id="benchd-eval",
                role=turn["role"],
                content=turn["content"],
            )

    def recall(self, query: str) -> str:
        results = self._memory.recall(
            agent_id="benchd-eval",
            query=query,
        )
        return "\n".join(results)

Quick start

Install

bash

pip install benchd-harness

Validate & test