mengram 0.0.11

Prospective + episodic memory engine for LLM agents

mengram

mengram is a lightweight memory engine for AI agents.

It gives you:

• A local, persistent store for “memories” (facts, events, rules) on top of SQLite.
• Hybrid lexical + vector recall.
• Prospective memory rules that react to events (e.g. repeated tool errors).
• An LLM-driven auto-ingest pipeline to extract semantic/episodic memories from interaction history.
• An optional FastAPI server if you want to expose everything over HTTP.

You can use it purely as a Python library, or run it as a local service.

---

Quickstart (Python library)

Install:

pip install mengram

Initialize the schema once per database and start using the client:

from mengram import MemoryClient, init_memory_os_schema

# Initialize the DB schema (safe to call multiple times)
init_memory_os_schema()
client = MemoryClient()

# Store an episodic memory
memory = client.remember(
    content="Talked to Alice about refund policy.",
    type="episodic",
    scope="session",
    entity_id="sess-123",
    tags=["support", "refund"],
)

# Recall related memories (hybrid lexical + vector search)
results = client.recall(
    query="refund policy",
    scope="session",
    entity_id="sess-123",
)

# Define a prospective rule (e.g. repeated tool errors)
rule = client.create_rule(
    condition={
        "event_type": "tool:error",
        "tool_name": "node_forecast",
        "window_minutes": 10,
        "threshold_count": 3,
    },
    actions={
        "actions": [
            {
                "type": "notify",
                "channel": "stdout",
                "target": "#ops",
                "message": "node_forecast failed 3 times in 10 minutes.",
            },
            {
                "type": "inject_memory",
                "content": "node_forecast is unstable, consider fallback model.",
            },
        ]
    },
)

# Record an event (rules are evaluated and any actions are returned)
event_result = client.record_event(
    event_type="tool:error",
    tool_name="node_forecast",
    scope="session",
    entity_id="sess-123",
    payload={"error_code": "TIMEOUT"},
)

---

Auto-ingest with an LLM extractor (golden path)

mengram can automatically extract long-term memories from recent interactions using an LLM-driven extractor.

from mengram import (
    MemoryClient,
    init_memory_os_schema,
    Interaction,
    LLMMemoryExtractor,
    interactions_from_dicts,
)

# 1) Initialize database + client
init_memory_os_schema()
client = MemoryClient()

# 2) Build interactions (dicts → Interaction helper)
history = [
    {"role": "user", "content": "Hi, my name is Dhruv."},
    {"role": "assistant", "content": "Nice to meet you Dhruv!"},
    {"role": "user", "content": "I work in AI & Analytics at Capgemini."},
    {"role": "user", "content": "I prefer morning deliveries if possible."},
]
interactions = interactions_from_dicts(history)

# 3) Wire an LLM client (replace with your provider call)
def llm_client(
    prompt: str,
    model: str | None = None,
    temperature: float | None = None,
) -> str:
    """
    Call your LLM provider here and return the raw text response as a string.
    This function is responsible for talking to OpenAI/Anthropic/Bedrock/local, etc.
    """
    ...

extractor = LLMMemoryExtractor(
    llm_client=llm_client,
    model="gpt-4.1",      # optional, forwarded to your client
    max_memories=5,
    temperature=0.0,
)

# 4) Auto-ingest memories
stored = client.auto_ingest(
    interactions=interactions,
    extractor=extractor,
    scope="user",
    entity_id="dhruv",
    min_importance=0.0,
)

print("Stored memories:")
for m in stored:
    print("-", m.content, f"({m.type}, importance={getattr(m, 'importance', None)})")

---

Concepts

Memory types

mengram currently focuses on two core memory types:

• Semantic memory Stable facts or preferences that persist beyond a single session. Examples:

  • "User works in AI & Analytics at Capgemini."
  • "User prefers morning deliveries."

• Episodic memory Specific events that might matter later. Examples:

  • "On 2025-01-15 the user switched from plan A to plan B."
  • "We agreed to review the forecast model next week."

Additional types (e.g. procedural) can be layered on later, but semantic + episodic already cover many useful cases.

Auto-ingest pipeline

The LLM-powered auto-ingest path looks like:

Interaction → Extractor → MemoryCandidate → stored via MemoryClient.remember()

Where:

• Interaction describes a turn or event (role, content, timestamp, metadata).
• An Extractor is any callable: Callable[[list[Interaction]], list[MemoryCandidate]].
• MemoryCandidate is a proposed memory (content, type, importance, scope/entity, tags, metadata).
• MemoryClient.auto_ingest(...) takes these candidates, normalizes them, and persists them using your existing memory store.

LLM extractor

LLMMemoryExtractor is a reference implementation of an Extractor that uses an LLM you provide.

• It formats recent interactions as a small transcript.
• Sends a prompt asking the LLM to propose a small set of long-term memories.
• Enforces a JSON array output matching the MemoryCandidate schema.
• Parses the array into MemoryCandidate objects and passes them to auto_ingest.

Important: mengram never calls an LLM provider directly. You supply the llm_client(prompt: str, model: Optional[str], temperature: Optional[float]) -> str function that talks to OpenAI/Anthropic/Bedrock/local models, etc.

---

Optional: HTTP API server

If you prefer to talk to mengram over HTTP, you can run the bundled FastAPI server.

Install server dependencies

If you’re working from source:

pip install -r requirements.txt

(If you’ve defined an extra in your pyproject, you can also do something like pip install "mengram[server]" from PyPI.)

Run the API with Uvicorn

From the project root:

uvicorn app.main:app --reload

This will start a local service and create a SQLite database (memory.db) in the project root if it doesn’t already exist.

HTTP endpoints

The service exposes:

• GET /healthz Health probe.

• POST /v0/remember Store a memory with optional TTL and tags.

• GET /v0/recall Hybrid recall with vector + lexical scoring.

• POST /v0/reflect Naive episodic → semantic session summary.

• POST /v0/plan Store prospective-memory rules.

• POST /v0/forget Delete memories by id or policy.

• POST /v0/event Persist incoming events and synchronously fire rule actions (if any).

Prospective memory rules (V0)

Rules capture simple “pattern → action” contracts that /v0/event enforces:

POST /v0/plan
{
  "if": {
    "event_type": "tool:error",
    "tool_name": "node_forecast",
    "window_minutes": 10,
    "threshold_count": 3
  },
  "then": {
    "actions": [
      {
        "type": "notify",
        "channel": "slack",
        "target": "#ops",
        "message": "node_forecast is erroring frequently"
      },
      {
        "type": "inject_memory",
        "content": "Last 10 minutes: node_forecast erroring > 3 times."
      }
    ]
  }
}

Each time an agent calls POST /v0/event, the service:

1. Persists the event,
2. Counts recent matches against active rules, and
3. Returns any triggered actions in the response (so your orchestrator can notify humans or inject new context into the next turn).

---

Python client recap

All REST capabilities are also available via the in-process Python client:

from mengram import MemoryClient, init_memory_os_schema

init_memory_os_schema()  # safe to call multiple times
client = MemoryClient()

client.remember(
    content="met Alice",
    type="episodic",
    scope="session",
    entity_id="sess-42",
)

memories = client.recall(
    query="Alice",
    scope="session",
    entity_id="sess-42",
)

rule = client.create_rule(
    condition={
        "event_type": "tool:error",
        "window_minutes": 10,
        "threshold_count": 3,
    },
    actions={
        "actions": [
            {"type": "notify", "channel": "stdout", "message": "tool is failing"}
        ]
    },
)

client.record_event(
    event_type="tool:error",
    tool_name="search_tool",
    scope="session",
    entity_id="sess-42",
)

You can also run:

python scripts/smoke_client.py

for a quick end-to-end smoke test without starting the FastAPI server.

---

Custom embeddings & fake-embed smoke tests

MemoryClient accepts a custom embedding function, so you can plug in OpenAI, Bedrock, HuggingFace, or a fake vector generator for offline runs:

import numpy as np
from mengram import MemoryClient, init_memory_os_schema

init_memory_os_schema()

def fake_embed(_: str):
    return np.ones(384, dtype=np.float32)

client = MemoryClient(embed_fn=fake_embed)

The scripts/smoke_client.py script supports a fake embedding mode via the MEMORY_OS_FAKE_EMBED=1 environment variable, which avoids downloading the sentence-transformers model:

MEMORY_OS_FAKE_EMBED=1 python scripts/smoke_client.py

---

Debugging stored memories

For quick inspection of what’s actually stored in your DB, you can use the included debug script:

python examples/debug_list_memories.py --scope user --entity-id dhruv

This will initialize the schema (if needed), connect via MemoryClient, and print out memories for the given scope/entity in a human-readable format.