FAI engineeringProject· about 180 min· server sandbox

Project 5: Build a memory system for an agent

Episodic memory (events with time), semantic memory (consolidated facts with source and timestamp), retrieval for a question, handling of contradictions, and a long-term evaluation over scripted sessions showing that the agent gets better thanks to the memory.

Teaches: Semantic memory and consolidation

Requires: Episodic memory for agentsBM25 and keyword searchAgents — plan, act, observe

Theory

Episodes are raw data about what happened. Semantic facts are distilled from episodes (key → value) and the latest source wins in a conflict. Retrieval returns the most relevant episodes + facts. Long-term evaluation: in session N, ask about things that were said in sessions < N.

Sub-tasks

episodic memory — Memory.remember(session, text) stores an episode with a sequence number; Memory.recall(query, k) → the k most relevant episodes (BM25 or word overlap).
consolidation — Memory.consolidate() extracts facts from the Swedish patterns 'X är Y' (X is Y) / 'jag gillar Y' (I like Y) / 'mitt namn är Y' (my name is Y) → facts with (value, session, episode id); newer replaces older.
answer from memory — Memory.answer(question) → a string from the facts or the best episode; 'vet inte' (don't know) otherwise.
forgetting + privacy — Memory.forget(session) deletes episodes and facts from a session (the right to be forgotten).

Passes when: accuracy_with_memory >= 0.8

The starter code

runs in an isolated sandbox on the server

import re


def toks(s):
    return set(re.findall(r"[a-zåäö0-9]+", s.lower()))


class Memory:
    def __init__(self):
        self.episodes = []      # {"id", "session", "text"}
        self.facts = {}         # nyckel → {"value", "session", "episode"}

    def remember(self, session, text):
        # TODO: lägg till episod med id = löpnummer
        ...

    def recall(self, query, k=3):
        # TODO: rangordna episoder på ordöverlapp med query (toks), returnera k bästa (nyast först vid lika)
        ...

    def consolidate(self):
        # TODO: mönster: "mitt namn är X" → namn; "jag gillar X" → gillar; "min favoritfärg är X" → favoritfärg;
        #       "jag bor i X" → bor; senare episod ersätter tidigare
        ...

    def answer(self, question):
        # TODO: mappa frågor ("vad heter jag", "vad gillar jag", "vilken är min favoritfärg", "var bor jag") till fakta; annars bästa episod; annars "vet inte"
        ...

    def forget(self, session):
        # TODO: radera episoder och fakta från sessionen
        ...

You write the code; tests you cannot see decide whether it holds up. Create a free account to run the lab.

Try the diagnostic Create a free account

Expected results

With memory ≥ 80 % correct on questions about earlier sessions; without memory ≈ 0. A contradiction (the favourite colour changes) is resolved to the latest. forget(2) makes session 2's facts disappear.

Common mistakes

Consolidation overwrites with older facts (wrong order).
Retrieval returns episodes from deleted sessions.
'vet inte' (don't know) is missing → it makes things up.