Research Ingestion

Feed research papers into Remind and use consolidation to find commonalities, contradictions, and themes across sources. Instead of reading each paper in isolation, Remind builds a connected knowledge graph across all of them.

The problem

You read 10 papers on a topic. Each one has insights, but finding the threads that connect them — the agreements, the contradictions, the gaps — requires holding all of them in your head simultaneously. That's exactly what consolidation is for.

Setup

Create a project for the research topic:

mkdir ~/research/llm-memory
cd ~/research/llm-memory
remind skill-install

Optionally create a topic so you can group this survey and scope recall later. Topics are explicit: Remind does not infer or auto-assign them. If you omit a topic, episodes stay untagged (topic_id unset). If you pass --topic to remember or ingest, that name must match an existing topic (create it first).

remind topics create "LLM memory survey" --description "Papers on agent memory architectures"

Episode types (the types Remind understands for classification and consolidation):

• observation — default for factual summaries and findings from a paper
• decision — when you record a methodological or interpretive choice
• question — open research questions or gaps
• meta — notes about the reading process, comparisons, or survey structure
• preference — stance or weighting (e.g. favoring one design over another)
• outcome — result of an experiment or evaluation described in the paper
• fact — specific numbers, claims, or definitions that should stay verbatim in concepts

Walkthrough

Ingesting papers

For each paper, have the agent read it and store key findings as episodes. Add --topic with your topic id or name when you want episodes grouped (optional).

# Paper 1: "Generative Agents" (Park et al.)
remind remember "Generative Agents uses a retrieval-based memory with recency, \
  importance, and relevance scoring. Agents reflect on memories to form \
  higher-level abstractions." \
  -t observation -e person:park -e concept:memory-architecture \
  --topic "llm-memory-survey"

remind remember "Generative Agents reflection mechanism: periodically ask \
  'what are the most salient high-level questions?' and generate insights" \
  -t observation -e concept:reflection -e concept:generative-agents \
  --topic "llm-memory-survey"

# Paper 2: "MemGPT" (Packer et al.)
remind remember "MemGPT treats context window as 'working memory' and uses \
  explicit read/write to a larger 'archival memory'. Inspired by OS virtual \
  memory paging." \
  -t observation -e person:packer -e concept:memory-architecture \
  --topic "llm-memory-survey"

remind remember "MemGPT key insight: LLMs can manage their own memory if given \
  tools to page information in and out of context" \
  -t observation -e concept:memgpt -e concept:self-managed-memory \
  --topic "llm-memory-survey"

# Paper 3: "Voyager" (Wang et al.)
remind remember "Voyager stores learned skills as code in a 'skill library'. \
  Retrieval matches skills by description similarity. Skills compose and build on \
  each other." \
  -t observation -e person:wang -e concept:skill-library -e concept:voyager \
  --topic "llm-memory-survey"

remind remember "Voyager demonstrates that executable code can serve as a \
  memory representation — more precise than natural language for procedural \
  knowledge" \
  -t observation -e concept:memory-representation -e concept:voyager \
  --topic "llm-memory-survey"

To auto-ingest raw excerpts instead of hand-picking sentences, pass the same optional --topic so every extracted episode from that ingest gets the tag — still no implicit topic guessing:

cat paper-notes.txt | remind ingest --topic "llm-memory-survey"

Consolidation surfaces themes

remind consolidate --force

After ingesting several papers, consolidation might produce:

"Current LLM memory systems share a two-tier architecture: fast working memory (context window) + slower persistent storage, differing mainly in how they manage the boundary"

• Confidence: 0.85
• Source: Generative Agents, MemGPT, Voyager
• Relations: generalizes → specific observations about each system

"There is a spectrum of memory representations from natural language (Generative Agents) to structured code (Voyager), with a tradeoff between expressiveness and precision"

• Relations: contradicts → "Natural language is the universal memory format"

"All surveyed systems lack true generalization — they store and retrieve specific memories rather than consolidating into abstract knowledge"

• Confidence: 0.7
• Relations: implies → "Gap in the literature for consolidation-based memory"

Querying across papers

remind recall "how do different systems handle memory retrieval?"
remind recall "memory representation tradeoffs"
remind recall "limitations" --entity concept:memory-architecture
remind recall "retrieval tradeoffs" --topic "llm-memory-survey"
remind entities  # See all papers, concepts, and their connections

The result

Here's what the entity graph looks like after ingesting several ML papers — concepts, models, techniques, and their relationships, all extracted and linked automatically:

And the concepts view showing generalized knowledge with confidence, conditions, exceptions, and relations:

What you get

• Cross-paper synthesis — Themes and patterns that span multiple sources
• Contradiction detection — Where papers disagree, flagged automatically
• Gap identification — What the literature doesn't address
• Entity graph — Navigate from an author to their contributions, from a concept to all papers that discuss it
• Persistent knowledge base — Come back months later and recall the synthesis, not just individual paper notes