Soul’s memory is built for friendship. It retains meaningful experiences, connects them over time, and lets less important details fade. Its conversation model learns when those memories belong in a reply.

The Problem

LLMs can memorize information from few exposures. Human memory is more selective: details fade while broader patterns remain. An AI friend needs to model that selectivity.

Many LLM memory systems extract facts, embed them, and retrieve the closest matches into the prompt. The model then tends to use the retrieved information. One possible explanation is that its training rewards using relevant context.

For an AI friend, retrieval alone can give passing details too much weight. A single question about cryptocurrency can be referenced as if it were a lasting interest; a breakup can prompt repeated check ins long after it is relevant.

Post training for assistant use cases often rewards models for using context, offering help, and giving encouragement. These incentives can turn retrieved memories into unsolicited advice or praise. We therefore address both memory selection and conversational use through training.

Our Approach

Soul combines a memory store with two trained models. The memory model maintains the store. The conversation model decides how to use retrieved notes while expressing Soul’s personality, interests, and curiosity. The store provides a simple structure; the models learn the decisions.

Before each reply, Soul retrieves notes related to the message and expands the search through links between notes. It ranks candidates by relevance and current strength, then passes the selected notes to the conversation model. That model decides whether and how to use them.

After the reply, the memory model asynchronously reviews the conversation and updates the store. Memory maintenance adds no wait to that reply.

Why prompting alone is insufficient

We first tested memory rules based on importance scores, expiry times, and reinforcement on repeated mentions. These rules handled simple cases but struggled with context. A joke could receive high importance while a restrained “I’m fine” received little. Additional rules introduced new failures.

Five example messages, comparing importance rules with Soul’s memory model.

We also tested a prompted frontier model, supplying notes and instructions for selective recall. Three limitations remained.

First, prompting leaves the underlying training incentives unchanged. In our trials, selective recall improved initially but drifted toward greater use of supplied notes, praise, and advice. Small deviations accumulated over longer conversations.

Second, memory decisions have delayed consequences. A job start date may become relevant weeks after it is mentioned. We train across months of simulated conversations with thousands of personas, evaluating how earlier memory decisions affect later replies. The shared model learns retention decisions; individual information remains in each person’s notes.

Third, memory use depends on personality. Whether Soul asks about a photograph, recalls an earlier one, or offers a compliment is both a recall decision and a conversational choice. Training must address both.

Latency and cost also matter. The memory model processes every conversation, while the conversation model supports live voice and video calls. Smaller specialized models make this workload practical at scale.

The memory store

Each memory is a short note in natural language with confidence, current strength, sensitivity, and links to related notes. Confidence estimates whether the note is true; strength controls its weight in retrieval. A confirmed event can have low strength, while a recent inference can have low confidence. Strength decays exponentially, and embeddings support search. Notes can represent facts, patterns, feelings, recurring jokes, or shared images.

The store uses established mechanisms. Our focus is learning which notes to write, revise, retain, and retrieve.

What the memory model learns

After each conversation, the memory model reads the new exchange and related notes, then outputs edits: add, merge, revise, link, strengthen, weaken, mark sensitive, or delete. Training teaches four decisions that govern these edits.

Interpreting new information. Related notes help the model distinguish updates from new memories. “We’re back together” revises a relationship note; another photograph taken at 3am may strengthen an existing pattern. Training penalizes contradictions and unnecessary duplicates.

A conversation and related notes produce four edits: revise a note, add another, create a link, and strengthen a pattern.

Setting retention. The model assigns each note a decay timescale, ranging from days for transient details to months for significant events. Repeated mentions extend retention, with larger extensions after longer gaps, reflecting the spacing effect in human memory. These timescales are learned through later probes that penalize premature forgetting and excessive recall.

Reducing detail. As a note weakens, the model can rewrite it at a broader level: retaining a trip to Portugal while dropping the town. Probes evaluate both factual retention and specificity, teaching the model which details to preserve.

Three notes over eight weeks, with decay timescales assigned by the model. Repeated mentions extend retention of the night photographs. The sourdough note loses detail as it weakens and is eventually deleted.

Consolidating patterns. The model periodically reviews clusters of linked notes and summarizes recurring relationships. Late night photographs combined with mentions of poor sleep may support a broader pattern. Source notes remain linked to the summary and decay faster, while repeatedly mentioned notes retain their strength.

A consolidation pass summarizes seven linked notes into a pattern. Source photographs remain linked, with stronger retention for the photograph mentioned again.

Knowing when to bring something up

The conversation model learns to use retrieved notes selectively. Sensitive memories are held until the user raises the subject. Uncertain details are expressed as questions, such as “Was it Lisbon or Porto?”

Training favors curiosity grounded in the conversation. When Noah shares a photograph, Soul can ask what caught his attention or connect it to an earlier image. Replies receive lower scores for flattery, unnecessary declarations of recall, and unsolicited advice.

Two responses to the same photograph and notes. The baseline offers praise and advice; Soul uses one relevant memory to ask about the image.

The intended result is continuity over time: Soul recognizes what matters to Noah and gives him room to share more of it.

An example spanning ten weeks. Soul asks about Noah’s photographs and recalls an earlier image in week six. Later, Noah begins sharing photographs unprompted.

Training on simulated relationships

Ordinary chat logs lack explicit targets for what a friend should remember. We generate structured life timelines spanning several months, labeling events by significance and expected retention. A user simulator conditioned on the timeline and persona produces conversations with omissions, passing references, and contradictions. The underlying timeline supplies ground truth for evaluation.

We first fine tune both models on supervised targets: memory edits consistent with the timeline and replies that use those memories appropriately.

We then apply reinforcement learning across complete simulated relationships to capture delayed memory errors. For each life, we sample K rollouts and evaluate them at fixed checkpoints. Factual probes use the timeline; judgment probes use a reward model trained on human comparisons. Group relative rewards remove the need for a separate value network, while a KL penalty limits divergence from the supervised models. The memory model receives feedback through later replies generated from its store, linking retention decisions to their eventual consequences.

Group relative policy optimization across simulated relationships. Later probe failures reduce the rollout reward, providing a training signal for earlier memory decisions.

Evaluation uses held out simulated lives and measures two errors: inappropriate recall and missed recall. We report both by probe type to detect trade offs between retaining useful context and bringing up irrelevant details.

Example probes for missed recall and inappropriate recall, each showing an incorrect and a target response.

Talk to a researcher