llama-index-memory-engram

Durable, explainable memory for LlamaIndex agents — powered by Engram.

EngramMemory is a BaseMemory implementation that replaces LlamaIndex's built-in chat-history buffer with Engram's hybrid retrieval pipeline (BM25 + vector + knowledge graph + reranker). Every message your agent sees is persisted to an Engram bucket; reads come back in chronological order, and semantic recall is one call away.

Install

pip install llama-index-memory-engram

Usage

import os
from llama_index.llms.openai import OpenAI
from llama_index.core.agent.workflow import FunctionAgent
from llama_index.memory.engram import EngramMemory

os.environ["ENGRAM_API_KEY"] = "eng_live_..."   # or pass api_key=... explicitly

memory = EngramMemory.from_defaults(
    bucket="user-42",         # one bucket per user / session / agent
    read_limit=50,            # how many recent messages get() returns
)

agent = FunctionAgent(
    llm=OpenAI("gpt-4o"),
    tools=[...],
)

# IMPORTANT: pass memory to .run(), NOT to FunctionAgent(...)
response = await agent.run(
    "What did we decide about the Q3 launch?",
    memory=memory,
)

⚠️ Pass memory to agent.run(), not the constructor

Modern LlamaIndex FunctionAgent.__init__ accepts arbitrary **kwargs but silently swallows memory= — your memory backend will appear configured but never actually be touched. You must pass memory to each agent.run(..., memory=memory) invocation. This is a LlamaIndex API quirk, not an issue with this package; we caught it in our e2e test and it's the single most common pitfall when wiring EngramMemory.

If you want it to feel like a constructor arg, wrap once:

async def chat(prompt: str) -> str:
    r = await agent.run(prompt, memory=memory)
    return str(r)

Get an API key at https://lumetra.io. Keys look like eng_live_....

Direct semantic recall

get() returns the most recent read_limit messages, which is what agents expect from chat history. When you want hybrid retrieval over the entire bucket, call query() directly:

result = memory.query("regulatory risks we discussed last quarter")
print(result["answer"])
print(result["memories_found"])

Bucket scoping

Pick a bucket name per logical conversation scope:

EngramMemory(bucket=f"user-{user_id}")           # per user
EngramMemory(bucket=f"session-{session_id}")     # per session
EngramMemory(bucket=f"agent-{agent_id}")         # per agent

Buckets are created on first write — no admin call needed.

Self-hosted Engram

EngramMemory(
    bucket="ops",
    base_url="https://engram.internal.example.com",
    api_key="...",
)

API reference

All methods have async equivalents (aput, aget, ...) inherited from BaseMemory; they currently run the sync implementation in a thread.

Configuration

License

MIT — see LICENSE.

For data-handling details see PRIVACY.md and https://lumetra.io/privacy.