Memory

Two backends, same three-method interface.

ConversationMemory

In-process FIFO window. Good for chat-style sessions inside a single Python process.

from agentforge.memory import ConversationMemory

mem = ConversationMemory(max_messages=64)
mem.add("user", "Hello")
mem.add("assistant", "Hi there!")
for msg in mem.get(limit=10):
    print(msg.role, msg.content)

PersistentMemory

SQLite-backed, session-keyed. Survives process restarts and is multi-user safe (each session id has its own history).

from agentforge.memory import PersistentMemory

mem = PersistentMemory("agentforge_memory.db", session="alice")
mem.add("user", "Remember I prefer terse answers", source="cli")
mem.add("assistant", "Got it.")

# Later, possibly in a different process:
hist = PersistentMemory("agentforge_memory.db", session="alice").get()

The schema is one table, messages(id, session, role, content, ts, extras_json), with an index on (session, ts). That is enough for ~10⁶ messages per session before query latency matters.

Plug in your own

Implement the Memory protocol — add, get, clear. Useful targets:

• Redis for ephemeral multi-process sessions
• Postgres + pgvector if you want semantic recall over the history
• A wrapper around ragforge.Pipeline so the agent can search its own past