AuthorsHanqi Jiang, Junhao Chen, Yi Pan et al.
TL;DR
SYNAPSE uses spreading activation over a Unified Episodic Semantic Graph to solve contextual tunneling, reaching 40.5 F1 on LoCoMo (+7.2 over A-Mem).
Read our summary here, or open the publisher PDF on the next tab.
THE PROBLEM
Standard RAG treats memory as static vectors, causing Contextual Isolation where relevance is tied only to semantic proximity and ignores structural links.
On LoCoMo’s ∼16K token dialogues, full-context agents like LoCoMo and MemGPT consume ∼16,910–16,977 tokens per query, leading to Contextual Tunneling and brittle multi-hop reasoning.
HOW IT WORKS
SYNAPSE centers on a Unified Episodic Semantic Graph, Cognitive Dynamics with Uncertainty Gating, and Triple Hybrid Retrieval to turn interaction logs into a structured, dynamic memory network.
You can think of SYNAPSE like a brain-inspired card catalog: episodic turns are cards, semantic concepts are categories, and spreading activation is the energy that lights up related shelves.
This KEY_MECHANISM lets SYNAPSE recover temporally and causally linked memories that share little lexical overlap, something a plain context window or pure vector store cannot do.
DIAGRAM
This diagram shows how SYNAPSE processes a query through dual triggers, spreading activation, lateral inhibition, and hybrid scoring to select memories.
DIAGRAM
This diagram shows how SYNAPSE is evaluated on LoCoMo, compared to baselines, and ablated by disabling key mechanisms.
PROCESS
Unified Episodic Semantic Graph
SYNAPSE constructs episodic nodes from turns and semantic nodes via concept extraction every N = 5 turns, linking them with temporal and abstraction edges.
Cognitive Dynamics Spreading Activation
Given a query, SYNAPSE initializes anchor nodes via lexical and semantic triggers, then runs three propagation steps with fan effect, temporal decay, and lateral inhibition.
Triple Signal Hybrid Retrieval
SYNAPSE scores nodes using semantic similarity, final activation a(T), and PageRank, retrieving Top k = 30 nodes as the compact, structurally relevant context.
Uncertainty Aware Rejection
SYNAPSE computes retrieval confidence Cret; if below τgate = 0.12 it triggers a negative acknowledgement, otherwise it verifies evidence via a strict prompting protocol.
KEY CONTRIBUTIONS
Unified Episodic Semantic Graph
SYNAPSE introduces a dual layer topology combining episodic interaction logs and semantic concepts, with sparsity constraints keeping |V| ≤ 10,000 and Top K = 15 edges per node.
Cognitive Dynamics with Uncertainty Gating
SYNAPSE implements spreading activation with fan effect, lateral inhibition over M = 7 nodes, and a calibrated τgate = 0.12 to reject hallucinations and adversarial queries.
SOTA Performance and Efficiency on LoCoMo
SYNAPSE reaches 40.5 weighted F1 on LoCoMo, +7.2 over A-Mem, while reducing token consumption by 95% compared to full-context methods using ∼16,910–16,977 tokens.
RESULTS
Weighted F1
40.5
+7.2 over A-Mem
Temporal F1
50.1
+4.2 over A-Mem
Multi Hop F1
35.7
+8.7 over A-Mem
Adversarial F1
96.6
+27.4 over LoCoMo
On the LoCoMo long-term conversational memory benchmark, which averages 16K tokens and tests five cognitive categories, SYNAPSE’s 40.5 weighted F1 shows robust multi-hop and temporal reasoning. The 96.6 adversarial F1 demonstrates that SYNAPSE’s uncertainty-aware gating sharply reduces hallucinations compared to LoCoMo’s 69.2 F1.
BENCHMARK
On the LoCoMo long-term conversational memory benchmark, which averages 16K tokens and tests five cognitive categories, SYNAPSE’s 40.5 weighted F1 shows robust multi-hop and temporal reasoning. The 96.6 adversarial F1 demonstrates that SYNAPSE’s uncertainty-aware gating sharply reduces hallucinations compared to LoCoMo’s 69.2 F1.
BENCHMARK
Weighted F1 on LoCoMo across methods, excluding the adversarial category as defined in Table 1.
BENCHMARK
Cost efficiency (F1 per $) for 1,000 queries on LoCoMo as reported in Table 4.
KEY INSIGHT
SYNAPSE achieves 40.5 weighted F1 while using only ∼814 tokens per query, a 95% reduction compared to full-context methods at ∼16,910–16,977 tokens.
This is surprising because many practitioners assume feeding more context always helps, yet SYNAPSE shows that selective, structured retrieval can be both more accurate and dramatically cheaper.
WHY IT MATTERS
SYNAPSE unlocks long-horizon, multi-session agents that can reason over weeks of interactions using a dynamic graph, not just a sliding window or flat vector store.
Builders can now deploy agents that maintain coherent narratives, handle adversarial queries safely, and scale memory without exploding token budgets or relying solely on brittle semantic similarity.
Nikhil Verma
· 2026
Focus Agent adds start_focus, complete_focus, a persistent Knowledge block, and an optimized Persistent Bash plus String-Replace Editor scaffold to actively compress context during long software-engineering tasks. On five hard SWE-bench Lite instances against a Baseline ReAct agent, Focus Agent achieves 22.7% token reduction (14.9M → 11.5M) while matching 3/5 = 60% task success.
Xingyu Lyu, Jianfeng He et al.
· 2026
ADAM combines Anchor extraction, Distribution estimation, Anchor selection, and Query generation to adaptively probe agent memory via an auxiliary generator and entropy based selection. On the EHRAgent benchmark with Llama2-7b-chat, ADAM reaches EQ=77 and ASR=1.00, compared to MEXTRA’s EQ=44 and ASR=0.89.
Shannan Yan, Jingchen Ni et al.
· 2026
AdaMem organizes dialogue history into Working Memory, Episodic Memory, Persona Memory, and Graph Memory coordinated by a Memory Agent, Research Agent, and Working Agent. On LoCoMo with GPT-4.1-mini, AdaMem achieves 44.65 F1 overall, beating the best baseline LangMem at 41.76 F1 by +2.89.
Answers use this explainer on Memory Papers.
Checking…